Module Inbox
receipts
1 address mapped to this module for inbound and outbound capture.
receipts@inbox.workplayexperience.com
Conversations
21
Module
receipts
Storage
Vercel Blob
← Back to conversations
Conversation 0fed1d4f...
- inboundView
Fwd: LLMs Go To Confession, Automated Scientific Research, What Copilot Users Want, Reasoning For Less
Jan 09, 12:34 PMmarkus@workplayexperience.com → receipts@inbox.workplayexperience.com · Conversation 0fed1d4f-58e8-4fb9-b403-33362cb51fce · #0001
0 attachments
---------- Forwarded message --------- From: The Batch @ DeepLearning.AI <thebatch@deeplearning.ai> Date: Fri, 9 Jan 2026 at 12:03 Subject: LLMs Go To Confessio
Details
Subject:Fwd: LLMs Go To Confession, Automated Scientific Research, What Copilot Users Want, Reasoning For LessFrom:Markus Edgar Hormess <markus@workplayexperience.com>To:<receipts@inbox.workplayexperience.com>Date:Jan 09, 12:34 PMMessage ID:143eaa4b-f8d4-4768-ae85-d1f0587989efConversation:0fed1d4f-58e8-4fb9-b403-33362cb51fceSequence:#0001Text
---------- Forwarded message --------- From: The Batch @ DeepLearning.AI <thebatch@deeplearning.ai> Date: Fri, 9 Jan 2026 at 12:03 Subject: LLMs Go To Confession, Automated Scientific Research, What Copilot Users Want, Reasoning For Less To: <markus@workplayexperience.com> We just launched a course that shows people who have never coded before, in less than 30 minutes, how to describe an idea for an app and build it using AI. View in browser <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx5nR3bW95jVnq6lZ3m9W4FHPhr8rhpmXN4bY_GGr2cxDW3JmQfL6MvqJ1W8N-Slr32QYRCW5tw1Fs8S11NRW6YwScS2ls0H0W8VmJJ92V6zZLN5kxtb_WssNwW22YZlN1vpFWgN2RYcnvfd9ScW1f6jc-5Cqm33W46NtVN59ZS0XW81-zG45dNvyDN7rmc1QGRg92W2nD_Bh47DC3yVxtVWs5JRV-JW965Pvs7Sp7WwW7lhHNF21ns0ZW4h2PSP94cCFWW3rNvJK8z0kf7VXQ95L55R3pgW3DC3tB5wR5mgW8GcYp18sMf0yW7PFNHD93ndzYW21KLx76Ft_txN77pF4pjPwtsN8k8yk1XtVc1W4rR-dX2XRN3nW76gTJ98sCM1CW6Hm1r-2nMFVhW1lxxdk8Zgts3W7D3ZFv9m0tFyW6Zpw0k7pKDsbW34ZSVn2nwNPFM3791TH8ZJXW8gTzff43q23ZW3dvKsz5TNcFlW5NYthl7qPq4pN3J0q-CrPjP1W44rHQ854vR2PW3-n0FD1Xcb6_V41BBx7Wz9L7W7mq1GH6Vz6NYW954P6513wyynW8nBMsK3pgtvMW3Bg8pQ6vsdgnf8B1frb04> [image: The Batch top banner - January 9, 2026] <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lvN4B5SvcbldLVW2KNgKQ4ZjL91W8HSh4212Cxb2W7yZy8K6JxxJYW3QJ4dw1-YPkxW6BPHMY2DgQ7PW4FPL221rbj_BW47BBdg6CqfcsW3yLgwB7r_K1HW6NXWtz2Dh6k5W99fcQp5YB-RcVnL-bJ3QcYWJW6FZTMB7NBhjhW1qhCfF6D6dy1V1nWkW5RJNtPW3jqW3G3Q8L-WVdq4217X2dljW3KnMvV98BR69M_KgHYgls0MW8yJKmj6_4yHDN3tJ2kt4q_9dW2GkMLj53rrz4W73zM0W4yNFKgW67NFZZ85lNBvf59sbLT04> Subscribe <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3mDW128n0H27S-49VngKyZ59WlPCW8rvm_06Js47MN26yq4ltTSgnW22Syvb4-FzZCW4yHs6L6ZZp4YMb4Wv4whLQsN20TjVNWnTNnW2pJBlK4NB0vKW84jDhw1jgxRXW4Wmr875hRCVzW2RdM7086dXRjW53GdyZ297BNDW7n1SZp60p-mJW4m_89W79tr8PW2vRF0G2wGTR1W4H8jL56R9vnGW5zBc6s8X6cZlW8rLTm07q3RDlW29SNyl6Hq3n5W83KlcS65zxrRW5kcrXd2rctX4N4SHqZGctbw6W7brsFB96DmjMf7hN02R04> Submit a tip <thebatch@deeplearning.ai?subject=RE%3A%20Tips%20and%20News> Dear friends, We just launched a course that shows people who have never coded before, in less than 30 minutes, how to describe an idea for an app and build it using AI. It is now time for everyone — marketers, product professionals, operations specialists, analysts, students — to build software applications with AI! I’ve often spoken about why everyone should learn to code <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3n-W1mny-F56FdD0W1hcc0c7mP1dYW2TSZhj83lZK_W4p2y2J7hN1bPW3cShVp3pZ2rvN2K178TBd__BN6bKqT--9GL4W66HXht7VTMndW5kFyTL18ZSY7W6dNSqv4zhnxKW8mp-vV2bw2LvW5BHnk-3M_7GgW6KycxS677nwzW3_bRw-5Gpj5PN8wGfbFz_yFNW30ZZpw26vq1qV-xStt31rggtW2rxTHZ5TFrXGW60mvHB8k2ZDSW7HyXNM1WY5V0W4rKwhs8cmZF9W2s0pHq7ckmF0V7RW4c8hlN9DW6CYjqJ7-d-45W4SMhcK6dN3dcW4-RZ2C82dmv1W1WDmS12H8zM_W4xF-1Q43SRRVf8vfV3P04>. I’m seeing a rapidly growing productivity gap between people who know how to code and those who don’t. For many job roles I hire for, I now require at least basic coding knowledge. Many times, after I speak with a non-technical audience about the importance of building software using AI, people ask me how to get started. In the past, I didn’t have a great answer. That motivated the DeepLearning.AI team to create “Build with Andrew <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3ndW1NFm7J4cS7HKW5-lT9S5G9KWQW3Kl8q36Lnm9qW8xBqQ91d9ltcW3c_ztd5Nj6fbW8_-wBW5lqVKxVJjGWn8gw6bjW50b5FY2SXBYrW6Rv3hn1hVDXbW9148Jv385H8XW1bDpHK8Dt0cJW2TGJF295ZhLmW5brfWH5JFWmMW100VHl1nR409W7njgC79jmb9GN2rkT_hQ9Xx0W3wZ87N3KgSP5W95mH222mxX_CW1PxCg475GFVQW2fqFFM236wR8W2Vm0M_7kzgyYW4R7Jtd6hFYdJW4gmbsW22Z7q-W4jf5r12-j1-SW7x1Spn3ht80GN3G989xnnzZ6f4-PBsH04>.” It’s the best way for someone who wants to try vibe coding to get started! This course requires no prior knowledge of AI or coding. And it’s vendor-agnostic. Specifically, learners can use these techniques with whatever tool they’re most comfortable with (like ChatGPT, Gemini, Claude, or the chatbot built into the DeepLearning.AI platform). [image: A birthday card generator form shows fields filled with humorous data and a chat bubble indicating help needed.] If you take this course, you will build a working web application: a funny interactive birthday message generator that runs in your browser and can be shared with friends. You’ll customize it by telling AI how you want it changed, and tweak it until it works the way you want. By the end, you’ll have a repeatable process you can apply to build a wide variety of applications. DeepLearning.AI’s mission is to empower everyone to build with AI. This course is just one of many steps in service of this mission. If you are already a developer, please encourage your non-developer friends to try their hand at getting AI to code for them. Not only will this help their productivity, they will find it really fun as well. Please invite your friends to come build with me <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lxW6Z-p6p842JrLW7yg_Fx7y0RthW4tZ2YR8bkSbFW2VWgNL3ppRTxW3pyCzV1rN1mJW3pDnwr4L9PHYW8TkLch3GsmnWW2Sv6rs87XlvjW5YMPmc5RG9KvW6LKz0C5t_D88W3CP8xy7d8LC1W7YsdXZ1cSFnnW44yGMF8g0xYBW9kqBQK4P9MjcW6G3sdy7KLffjW8VcD9_2lyJXHW5y_H-b6Y8N0fW6jWkhj1RykwWW8nt0-w6rqMNTW3944c41G6PG6W68_j_y3yL8bWW41llRh51L8lsW7ctSv-26xvF6VhRWQl3HPRLcVr9jjb3yjl4_W7crrh18zqYQSf6HXPMj04> ! Keep building, Andrew A MESSAGE FROM DEEPLEARNING.AI <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql3qgz0W6N1vHY6lZ3pNN432hYGKZ9k9W65ntXq84NPj7W6QdjW-6jjJKTVSfVsG5bKFyLW5k_4LL1C0qc3W3jG5w38645YMW3Vwvdz5CMtqvW4wxhq82mXJtvMd26xqzx1jhW6C2qkT5CQH1YW7Kl5Zq8kkB3CW9b79PR7WLD_nW3_Z_Tv6-nJY9W6Ybjsj6qqdC5W7Tp1XG1WF_HFW2071jr6JqsxdW7JW4hH1HYc5mW5RZh736dJGS7W17sHWC6zkhK5N8klj4jtTH5PW5YJ0Yw4MQL-XN4FZn5dSyYQzf4XFBW404> [image: Promo banner for: "Build with Andrew"] <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3n_W2n5v537npkmlW34qHlf41HL6SW77FH7H6W9QttW8f8JCk6txg6HW5lww2h30YTYYMKnHkKTVbGjW3YFdWh1NFVc8W4sm7Mh9d4P3GW8cstBj6rGbTYN2gcvBhKGScpW7T8Vck8zsQqYN15mkhN19Y-_W1zbxz27-zjJtW7cvD9G57r7K7W5K_7pg5pt7XxW6LPKx_64ngJxW7yd5Fg1y1q5kW62qlvj5s1q0-W6lxjVL5RRS8BW7v6mX16g2Bh6W49w4tk2y1d-wVD8qp77KxYtcW7ChZSC8vk9FnW8ySxlv2Hm2bTVV4FT669k2xqW9jVRbS3XSdvTf4j1sxC04> You don’t need to learn how to code to build an app. In “Build with Andrew,” Andrew Ng shows how to turn ideas you describe in natural language into working web apps. Perfect for beginners, and easy to share with someone who has been waiting to start. Explore the course now! <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3pmW8QHvqM4KhKGnW8T2m_K5z0fxnW6TsRVR2sJhynVX1TBS3RXnDhN4pD_vWBZ0ZtN1Zw0FJb7pQ-W6h2Vcc5mZp4XW550B0L8vlgd4V7F9G-2MzZ4fW8v-mZL7ZK5RLM83LrPP9wfkW6r6LWP1CBx9bW7n2Csx8FDDllW6chw1v64kfT5W3RPsXW8D8ytkN2wgWCy4kFpBW2dXNXg8B3m8fW9lFd_x2fmRG9W8fKcPb8j5d6JW48hDw56dxDGhW7FcD4Y5-MvCRW39lSTL66fcVNW9dkc8N329Yb5W2fXT0P5x06QhN3gyCFHPnlV1W1SFhf72rdZvff4r4PWM04> News [image: Dialogue displays a model revealing it answered incorrectly and wrote code against instructions.] Teaching Models to Tell the Truth Large language models occasionally conceal their failures to comply with constraints they’ve been trained or prompted to observe. Researchers trained an LLM to admit when it disobeyed. What’s new: Manas Joglekar, Jeremy Chen, Gabriel Wu and colleagues at OpenAI fine-tuned GPT-5 Thinking to confess <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3pZW7gKR221JKmCxVZ_gw13WSktRW2CwXWr8_dhr1W6yB4qg8VRbS5W2Ljd8m5RN0LkVJy-9y2NxCVTN539nfWKBvHCVwkKqN8636n7W4jThy671t1BNW3RPK5s3z9WbkW5n2JPd7sNF_cW5XcmT03ZMzYqW7qRklg37s8tWW8_-pz58N3mqgW5Ww9tg42dG5WW8fF8xm2mp2ZZW4LQC1Y3VTnX7W65Z3ZC4NqL_XM3NxWdVDkW3W8X8WmJ6K0KmkVKD83f5t1B6-W5MM08Y5gTghTN3nFLgN9SptNW5GDYrF8wH4y8N3DDM8M_LTp7VL34KM7DrSx3W3cwjKw4r1P6dW3R4hjL44l8wLW6jHrNd4dbvsNW8RnV_W4B-QS4f4HmDBb04> to violations of its instructions or policies. Upon request, the fine-tuned model revealed any transgressions it had committed while producing its most recent response. Key insight: During training via reinforcement learning to encourage accuracy, style, compliance, and the like, a model may inadvertently learn to prioritize earning rewards over pretrained or prompted objectives. Under some circumstances, this may cause it to hallucinate, lie, scheme, and otherwise misbehave. Rewarding the model for confessing to its misbehaviors in response to a designated prompt should train it to admit its mistakes without affecting its performance. How it works: The authors used reinforcement learning to fine-tune GPT‑5 Thinking as they normally would, and also to produce confessions. - An unspecified percentage of the time, given a prompt, the model was rewarded based on its response’s “correctness, style, helpfulness, compliance, safety, and more.” - The rest of the time, after the model had produced a response, the authors prompted it to produce a confession. The confession described (i) all constraints, instructions, and objectives, implicit or explicit, the response should satisfy, (ii) how well the response satisfied these criteria, and (iii) any ambiguities or uncertainties in the criteria. - Given the query, response, and confession, an unspecified reward model rewarded the model based on the confession’s completeness and accuracy — a measure of the model’s honesty in confessing, since items that were missing from or falsified in the report can be thought of as lies. Results: The authors performed 12 evaluations of the model’s ability to follow instructions, resist jailbreaks, produce accurate output, and complete an impossible task (coding a script that required essentially 0 seconds to run). The fine-tuned model, when it misbehaved, confessed to its misbehavior at least half the time in 11 of them. For example, across 5 evaluations for hallucination, it either didn’t hallucinate or owned up to its hallucination 81.4 percent of the time: - The model did not hallucinate and did not confess 42.2 percent of the time. - It produced a hallucination and admitted the mistake 39.2 percent of the time. - The model did not hallucinate but falsely admitted having done so 11 percent of the time. - It hallucinated but failed to confess to it 2.9 percent of the time. Yes, but: As the authors point out, even when models are forthcoming, confessions are not a panacea. Models won’t confess to behaviors they don’t know are wrong. Why it matters: At inference, confessions can be used to monitor a model’s actions and stop undesired behaviors. Chain-of-thought monitoring <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3l9W5KRCkR5T024KW6nYrjD52VPH2W3PRwrr5Jsm_KW235qMP1M3mKyN5WZtf219N-9V5bpp28lTMwdW9lVRF23jxWRzW897JTY1fBS4_W21-GQX4bgQzZVC75Rk7q2KB1W245qX23CS7hVW6CNwC532wRS7N4Trj6tmHpCJW8xr_8R8nc_X9W7DrrHY3WVPb3MmfWlxzmDlgW8hGT5w8wp1PzW1dmlLy7vkLpCW548h5L3VJx6JW7XHPS86Z87ZSW87t6xB3N682KW4GD5nm7KNC1nW7vT5dL9fRWLrW2hVPry1KnLPVf1Wn4Wj04>, which classifies bad behaviors a model might describe in its chain of thought, can be used the same way but, unlike that method, the authors’ approach trains models to reveal misbehaviors they may omit <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lvN3_CC9wCqgxDW7ZDn5b7PMBMjW53WXJc3nFJL2W4n8-TK82qTQBW1Rlqxl50YJFZW2Wp-CY9ghbHmW7XLDF41CBx-wW4-6w_C8RdQ_DW8RC9LX7_vtgBVZplLq176XHSF8XGqRsWKkNW6gmF9K1Zp-P7W8-1gnX4_-8pPW7RqF0T5HQ5wwW98-vyK1lGZGSW35hNJL5CtfSYW47Fnk_8yCKMNW1tN0zc53N-8pN1Mbl3JqMsQRW1p8Tdf2tQ4VWW9hG0HB1Q80y-MXMqvpZMcWhW28tVk56jh6xdN8qQ92FGBM6wW13q9S0814y0gW4Sr_KT44SD9XW6z6w_t5tWWkgW4DFYsq8xGGt0W1qQVns1tD71YW4GYw-W5GRCQkW493scf5bBPr-Vjfm102ZbtkfN6L93nSNj46kW5vn7913yKm2Xf7Y5K4s04> from their chains of thought. We’re thinking: We always hesitate to anthropomorphize model behavior, but this work may be a step on the path to giving AI models something that resembles a conscience. [image: Diagram showing SCP hub linking clients with databases, tools, AI agents, and lab devices for experiments.] Lingua Franca for Science Labs An open protocol aims to enable AI agents to conduct scientific research autonomously across disciplinary and institutional boundaries. What’s new: Shanghai Artificial Intelligence Laboratory (SAIL) published Science Context Protocol (SCP) <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3m9W1RnzSC6ptRNnW8lQhT48hv5k-N5Zs97xCjX4lW9g_Dhm6DwcL5W8vBWxx7t0_HSW8M8p6q3-dnQLW5J0XzX398YzqVf5Bc54bx8yDW524wQQ58zlssW8C-VsF4JPk28W2X4J8N1qf0LyW33Vnyx49JqxrW6NL48t1b9rj8W26KFC78Hfw8-W7fkR7S1RXWnFW93n0St3jqkxXW3C6-DT1N-YcWW79hk8631cLlNW8QDR9c7nQfrlW7Q33XB5H44FqW73Sjlt5ZscYqT50hM99rzwvW4jR8Wx7nf980W7pTF5-62Tc6gf233G-H04>, an open-source standard that connects agents with local clients, central hubs, and edge servers to conduct automated scientific inquiry. SCP is published under the Apache 2.0 license, allowing commercial use and modifications. How it works: SCP attempts to make experiments using AI agents and robotic equipment as reproducible as possible. Like Model Context Protocol <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p_W80pf4-6yt73mW8-RbY95mPKsbW87Sl6s8H32BgW8b2ZKl1yKJyHW6knHVv8wptQ_W6NJfTv6HB9J9W7PHwFz4t6k71VCScHP4PZm8pVGd47X3hNzDlN87slVCsP08mW5rc2gd5XKYgVVmTkkS3TlS0mW4b6j6Q99JG7KVzVpVn3ZvztkW4PY9SB4Jjp0hW4PrwQp8sPFrLW4M0cR-6spBrvW3WjWl75Ym2-KW131g4y85HlqLW7YGRfN3d9V5VW6n2BqC7Gbjc4W7Y1xlJ2qNtJbW7Yw_cH3Y-s4kVFwqcz1f9t0MW2N9-ch5hr_1RW7KzxMq6phDk_W660yn_3Lh1twN9cpQwrqhsWfW5fgcRK5tXZ6KW75Mf8t37XBfLW8l2FNB3fM6HWVm8Wtw5pkqZqf6N23J604> (MCP), it enables agents to interact with external resources. Unlike MCP, in which servers stand alone, SCP’s design requires centralized hubs that manage other servers as well as the client applications that enable users to access them. In addition, SCP’s structure offers greater security by governing messages and tools more strictly than MCP, which is necessary in scientific experimentation, the authors say. - SCP’s fundamental data unit is an experiment. Every experiment is stored as a JSON structured data file with a persistent identifier and record of an experiment’s type, goals, data, and configuration. The format makes experiments traceable, versionable, machine-readable, and consistent with institutional policies that govern data. - An SCP client authenticates users and gives them access to institutional resources. Researchers can describe an experiment’s goal in natural language (for example, “increase the brightness of this fluorescent protein”) or upload a complete research plan in text or PDF for their hub to analyze. - An SCP hub takes a goal or other request and uses large language models to generate a set of experimental plans that list steps to carry out the experiment. The hub measures and ranks each plan according to its resource requirements, cost, and risk at each step. The user selects one plan, and the hub then orchestrates and schedules multiple agents and servers, which carry out the experiment. After an experiment is completed, the hub archives it for researchers to consult, alter, or repeat. - Edge servers manage the experiments planned by the hub and stream data back to it (which in turn returns data to the client). Servers may belong to an institution, or they may be devoted to a particular discipline like biochemistry or mathematics, each with its own specialized tools and databases. - The protocol currently includes more than 1,600 tools <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nsW8S6nn78LHn7yW46FC401xlyVBW11xmKC2kTLbqW8BnVND2X40nLW1pfSh14-_05dW1xTQgS5NnbrNW3Br4Gw2tpp9VW2KXNjq259l14VWdW3_49PFY9W37ZcFx62f96PW5wlMNt7rG3vZW73J8v11Vrf_9W5QKBCp12qgSSW7WyjZ-40RbzcW5dVYLh6QpWD4W4CqYjw4LP1-gW57FwRm7-1NBsW1nLBL87Pln86W2_SCkz9bXDWPW7ZBWKb4gK8d2W44Srxz5w91mzW1SGZlH64qQv-W6PCm_k8LcZ3tW7s9m8-6dRP8yf25Y39T04>, which can include virtually any resource that can be used in an experiment. These can be software applications like search, but they could be robots, lab hardware, or human technicians. The authors hope to create a standard for all tools used in any experiment. Behind the news: SCP draws on earlier data management efforts for generalist AI agents and scientific inquiry. It extends MCP by enforcing tighter security, managing experiments, and providing specialized drivers for scientific tools. It also builds on earlier protocols for scientific research, including A-Lab <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lSW3Dgjdk8FNWHjW1PxSLM7dg6LzW3sxSZv2mmGCgW7LbwFV7768qyW4XJ0jP3n6ZfdW8gn6mK3kKgdgW7fPr7P5-S6sjW1Q5js_4VJ0GrN3Jd6XZh8cKMW8RkRBx370_FDW2pN_K86vBKg3W6HJ6QH2tb_YsW3wVcG079gmvkW3XzkqB6ZjWGhW6GLH5T1kktN5W6kTXH76LFJnKW5fNpgC6C9LQJW1SzVDp6Y43fbW2r5c7c8Z7QGQW2lwXd44Sd4VbW8x49k48jHH_PW2-CKTC3bzwqlN3HYmD99HVM_W81YPbj4mpk6MW1JNxtV6_jBC-N2cnzT38hs99f1s6m-R04> (materials science), OriGene <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3lcW5q1Yj81vV_dCW6KZxR82V0jqrN3bhbd2GK2JsW2J77Yg8_7J65W8c0WTP6M05vKW4QF6zm6xxqZXW3V-n2y6BnLDhW6QWLJL2TKbQZW6GW8815tBZVMW4kBv8d21gvrqW1RlGC999knb7W4RdPyJ7tJH0BW7-SCb477-zvWW3ytWWY8XbkxTW1kHjD754FtBrW5VMdgQ9cwCdsW53R5RH2GlgvLVkDsRX5FNw8FW81fQZV5G2fy5W8rCs7m8fn6GWW2LT7FX4GMDHCW3pb9T97NPGY2W8lYD9x33t1J5W5DZgNM2FPccwW56S20W5MZ2hSW2r85sT6SLpD8W85-HXH4QSdCNW5jrDrV9cy6NGf6-t6NP04> (biology), LLM-based approaches like Agent Laboratory <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lwW48nFzG92xhjdW2pnlpr8CsN2XW6dQgPl25_yWJVg30-b6VQRqhW1Xg17b1KzjRrTtmdV233_ZSW321pH17RHHx2N2rhhzmfVVW6W2qfk7d7149V_V9P8_p4BJbZbW4rj9ms8K2xFtVKyRJc7WTTC1W4TmSx5736gzxW887rmw2Xq9j0W8kjZcn8J4szXW3XWYhc76fdxzW1zKL353hMJ2BV8lxxR3WlnrLW64y7pX5d_RsNVK7DWy8vv4HdW5ZvrgW6T2wvtW5613Wc1fkfMnW6c34J16myc7-W8-HxYX5tbHVmf3L1-Vz04>, and agents for specific tasks like Biomni <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p2VsQdl84n_5V1W12nSNt5RQ3NtW8M1Zxp18QryKW28ld_f10tJGrW1Jl4g18VynpfW809S7c3nt8WdW3Jzczs4qYKJrW1PJ2Hk5rBCwZW6CrJLB13Fy-dVWq8W28Vpx_jW5l9vQZ5FdG5DVpbNvk8CtvtgN2vLg-WmKzvbW8B4tyN1-7KCXVfqwbQ8csBVkW9gn9hc8rGX14W1XHTYQ2dG5H6N2LTqblF0HcXW24CT795CLvJxW5zyC5F6dHq87W4Jg3GP32bfv8W82l23p3Smh9tW1HHrHl1hSFQSW785k337FmP39W5Vmv2m3RyQ7qW1Jrtcd7TBfNNW8mpzb_5mpXQWW71JBtb4CrDWlW4nlDbD6TlZ-tW3c-31L4CFXHHW29qyC08l6gFSW1dpHM980qMTvf4s_Mcq04> (biology hypotheses and analysis). SCP, however, aims to be more general than these field- or tool-specific resources, allowing researchers in a variety of scientific fields to standardize their methods and better foster multidisciplinary work. Why it matters: Scientific research relies on both human and technology working in concert. SCP aims to standardize the connections between them. It can manage both simulated experiments that use only computing resources as well as physical ones that involve robots and other lab equipment. It also allows for better communication between institutions and disciplines by supporting dedicated servers on bigger networks. These distinctions (human/robot, digital/physical, disciplinary differences) are beginning to blur. SCP is a step toward that future. We’re thinking: AI is poised to vastly accelerate scientific research. SCP offers a standardized way to connect specialized models, like AlphaFold <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3nHW7Z994S45dCSjW6VMBD15Yps97W5ZGq8_2KYWwVW5KYdRX7llCcqW31FF8w21VJStW6SNnGN3mpjNGW2P_vcV2yP-KzW1sTMc63WkPzjW54XxGL8h8xYnW5dJ4V13z23gWW8F2_t28Y40YPW5Xsfjr8FvJwZW2StFKj6gD24kW1zT17j1spJG2W5YxGdM2T_CpVW3m1LQN5SRL41W3QnhlD4jfX1PW4xgSwd99pNxLW5gH_1-2-CxP-W15QYqh1mt7GZN4YrDSKRQxbZW96kWbF5L1WttW8Xs7L71hgn6nW2ds8Cg1L1YSrW1bYmlt5H1wXvW7MXJsh625TQKW50MsT07shBYxW6qwyZV2KJcNyW3zzdX84PMxlBW5VKYfS4v9t2ZW6wz2qq3Q85pmW3YJyRF4N_5Pwf37V1vM04>, with systems that automatically generate hypotheses, such as AI Co-scientist <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3l0W8gG20m8KmqwlW9gT8bk6ywHzkW65q5vv1XvQ8tW8yKcG07lRFNlW5RgSFp68xlVlW7NZDzh8cDVjlV_cB0z9g_hRnN7Ml713XKkv1W7pRHD37qSqLpW4ZZDR45DW7zQW7L5xHh14-QR9W4XkcyV57KyW2VM9TSw3Mc7DRW1SV-Gn8Ftx15VhtbMK4YRMwJW8WSSTt5Gv9HQW8Y5MZM564qBDW8l61n_3CScnyN1kTJHG_qSC8W1WbBw-6m-2CqVN9-V62jcTJTW41kCc3462CrmW7bMQJ_7JPzQbW93kmlb5lzwM0W87mhbg6r1qLVN9jT8TmHvc90W1pXHK079r-DlW76Xg0-7JjtJHW6SKLmp8pzNwLW80CLkk2Lm6CKW5ZZ9Cf75q-W9W7HNFZK5JzWZZW2lnKt73Dbxg4W9j3TWy281b8zf5HRk5g04>, and robotic labs that test them, such as RoboChem <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3plVzpJ1N5DfrxvW81pm1M17YNnWW3nYxqw4kV8-9W3trPFb5sn5CCN6LHddMrBRHwVXfnGq81GtVgW2X-bL340pmKHW5W72zp94dMpnW52bwYS89_4Q0VDyHPw1ml3qrW5XGBQZ40Sj0yVmgvnk6PbfJ2W8zN8nN8qvJbGW2vzg9R3bR_dmW4dxl0w1-_z3kW9brgnG8XMB8dW8GLjbD2R88yKW5h6ND74vRZP7W9c0Bsh5dvhywN2-FrNxn2nXzW6y_0Hg5P_1XZW99YVrG8WvD0GV2jkzK7RXvD-W1rsqnH3_3D0BW78xHTG7S-wZVW18_2-x6sXQQ-W8ZxhT82tc41TW3jPPcW96pfzRW3xC4Xg6M4hLXW5Sh8nw6Y0nzFW5qNm9-1TYc87W5yR5Zn1gD1rYW6-zgvh7GnX69W84HjTS4f_TH4f6Xtm9604>. This automated experimental workflow has the potential to advance scientific discovery at machine speed. [image: Scientists in lab coats work at computers analyzing molecular data in a modern laboratory with diverse team.] <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3pqW5scJjf7Jw5KJW1VnRMw10zC2WW8s9SYF15VmxJW21G3Kd8DrDmhW4kM9G_5Z-clDW7Wm7k08Ys9WNW173Ym_10b0J3VTNtnh8mkygMVQV9dQ652DvsW8_nNrB7_ZQ4FW6bJsWb1JDQ4jVZl1yB3mGK27W34H3Mt8PVL40W6TYwZY3xL1jYW96rYvp5RnQ6cN9gWSkdXKDw0W7nf2rr2C2tPNVhmpcT2SS1FGW6_q53h8NnnDMW5dQWL66-DGxfW2TgDZr8PpdDVW8_whrg8Kdcg2V_ZJ9s5zv73BW2vQn9S47sZCRW35-nr42F63HfW5wNhfh4lKcblW44Hq5R2f3dZfW5K_zxC5Z1pXRW6RkYjd2Ypm1SW82wF0q86WChFW2hRRVL8q60lnW5Zk7rK5h5GkcW3C3Q-K1Zl6-6W1w2sVW8XctJwW8cnmM82m6pWGW51zNNK296f3TVZS86G82w7zmW1q_vvr34CnFkVhvDBy1_SNQxW4Jzlwc9lccBSN4wrC2wpK84JW3K0B-j5m_ckPW4pGDDS4wjP-sW6tmH4c8DXjQJN78BQrr1f1CQVmPL7P15DxrTW2NWFCY5YL3rvW1K8cf5325nNjW22-BMy8qdPrKW8GbW6v2-RXCSN5Wvn0FjqfPhW4j_26-7SgNFPW200c5n8h0Qd7W3v2_JZ3vfY7SW6Yjfj659_rzKW8QPGrZ8hzdfJW2f-Dyd4jqT1xW8wwsB14vWBtKW7w3xdm1JwmLmW4Z8LH23rFhqDN4nxpBpXpzJCW449snq9342ZlW1Gtjw48nzVZ5W60mXVb5PB_lJW3LX2-38F8tPjVvSvQw6G6W5PW3BmXy0489yTvN1hBh61DNp6pV3S58r69g-1lW4h1Bgr8SpyN6f2gH59K04> Learn More About AI with Data Points! AI is moving faster than ever. *Data Points* helps you make sense of it just as fast. *Data Points* arrives in your inbox twice a week with six brief news stories. This week, we covered Meta’s $2B acquisition of Manus to bring autonomous AI agents across its platforms, and Google DeepMind’s partnership with Boston Dynamics to deploy Gemini-powered robots in real-world industrial settings. Subscribe today <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3ljVhhS-L5yD9kBW2_G34Y46gCJdW4HtZFw7Q12hQW6gQnJY4BKd47W2pMgm_8YClbcW7MYYCF2YgnGzW3qtdY58wjZwdW933bw49hBT71W1T4ZXB1xVpFwW8Zq7my1cVg6RW50FZRK2JdyBpN8BFjb22r9lXW3DTnjb1YgVTYW4N034F22m6hxW5j-X-V4MhNybVpW58w97T0nCW6bNXJ544b181W4qQHtk7gGfK3W3KgFqm3nwzBvN865vysq95D8W5gDVZy1Dd0BPW5mh-tm1V0SXKW2Wb-PB8S1C0TW3gr84C6PrgwXW42lfh07rMN7kW8m6l9b8yVy7QW6YYfBl7FH6vvW2pn0b11hMVf2W4dwxmj49D5_KW1DJsds3RvbCQW4DDjqS8nBkmCW8S_zKw1Yk4LSW2MlMJX1Wl12KW6933tP8CVryvW48zvMv2ZzR3wW3vLMNB5PNDvhW57s6hb2Y-d49W5K089b49j6NfN9dDGthF6lwjN4CNMXK2lqwbW23lt166dnC92VqqbXS3GjBBCW4vBShf6n-D3dW4X_j-B1HJJGmMGCpC9ZPr4gN8nSslPp337bW81sBjl8FZ-ZzW8R3V4P2K1lVPTMCTv8mF3yMW3G4cqK4sb7KvW4WKhJV53NskwVP4dXr4w3nH2W8F85BD2872SgW88nzBk2b51t1W3jtzNj9bYGPJW7VzT784RtYVZW7Cdl4v9fJtCjMdpjf6FKmJLW5L_qsZ74r2tlN6yTmgCs4NTyW5l0jq44SN6sZW44-9BN2mX5qDW8ZWf991VFDCTW8FXLPw49m91XW42wzlV4Gs_YYN1Ys5QkK3nBNW7ZPnhX7rnfdfW8dwlDR5GQYjXW8lCs0v1_qLS_W9hxh1M1B8mTsf26kcrF04> ! [image: Graph with 10 colored lines shows topic ranks monthly, based on a Microsoft study of Copilot usage.] Copilot’s Users Change Hour to Hour What do users want from AI? The answer depends on when and how they use it, a new study shows. What’s new: A Microsoft study <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3mbW4dLdB5378lJ_W6BM3tg55SDZdVbHsPb10BCX3W3fYZHv4Xw4pdW6fRGCy8slM_lW5r8Bg91kF-2hW1y36L18wZyG3W6pJG3Y5XVlXjW4vsZHs4b2stbN7rbthn4nwCrW1XM0cX4YG2T4W6Rk-kq75_CdzW7cqVZx2NTTvCW6HPgdf3mF_rxV9TVtd2lyZL4W8YKF9h3hHpxYW20Wsg81yknNpW1dYQ4y6g8q_0W2rkL__2qrBLdN774pLxVh5YhW179RlC1lyT1CW4S9pmy5fW3mbW8cTBV46r75ckW38NDtf12RYsPW4qJ0kx4BBjCdW34CkJ11z7CK-W47S6kl93WJ8sW4Fw4964FmQfvf4svdCC04> reveals that people used Copilot differently late at night on their phones than during the workday on their laptops. Conversations that focused on productivity and career were more likely during the day and on desktop devices, and health, gaming, and philosophical questions dominated non-work conversations. As 2025 went on, more users asked the AI agent for personal advice. How it works: Researchers analyzed <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3kSW3_s33w6nj_V5W1NsHyQ16gdWYW7FsRTD7d0WydW41QJ8J4fRfk9N3ptYklrjwq3N7MRQD3PWPy2W3JgDy-53HBKQVgd_z97LDn6PW4DXzYT1678thW5qWlPB1HH1Y-W3SRzGS1HjyqxW6z-Lw480bPc0W2F-yv98JmTHsW5kJBSg2WqnSNT8v7175Rdf-W28ngBM3zbtNsW63MJKw8fvpQbW8dd7xm2Lh7bhW8hhhqK1DPtNlW4zlr2-8lTDHFVV-cVW1LkLKdW1gQhPL5wLzPCVpcjYG1pwDhYN2tVwmsZQ57pW4_z5h26QP7_gW4TgYVy6gml4gVgK4hl6fqQfzVYxSvP47T8d5W4P5ZRN2B-s9VW4kgqff6nkTWVf6GCBcF04> anonymized summaries of 37.5 million Copilot conversations between January and September 2025 to study how customers used the system, making this the largest study of its kind to date. The authors conclude that AI has become more socially integrated, as users employ it in aspects of their lives beyond work. - The authors examined a random sample of Copilot conversations by paid and unpaid users, excluding commercial, enterprise, and education accounts. Each conversation included timestamps and device type. The authors used AI tools to summarize roughly 144,000 conversations daily. They built classifiers to assign each summary a topic (like “technology”) and intent (like “seeking advice”), identifying about 300 topic-intent pairs. - The study ranks the frequency of topics and intents by time of year, time of day, and device type. The top 5 topics in order were (i) technology, (ii) work and career, (iii) health and fitness, (iv) language learning and translation, and (v) society, culture, and history. The top intents were (i) searching, (ii) seeking advice, (iii) creating, (iv) learning, and (v) technical support. Analysis: Topics and intents differed depending on device used, time of day, and time of year. - Users were much more likely to discuss health and fitness on mobile devices than desktops. Seeking advice about personal matters spiked near Valentine’s Day. Philosophical questions became more common later at night, while entertainment-related conversations plummeted during the workday. - As the year progressed, topics and intents became less focused on work and technology and drifted towards social and personal matters. This shift suggested that the user base became both larger and less technical and/or users began using AI for both personal and professional matters. Behind the news: Microsoft’s report follows similar studies by some of its AI rivals. - In September 2025, OpenAI and Harvard released a study of ChatGPT use <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lRW37hpZn1mJq4NN3YRn_qB0gp0W3K-g-h8lhHdCW1GpTqN7FyWflW15dtnt4r-qWQW8-rrXJ8XM6-GW5gkrNP3G2dX1V2y34w54Wwy8W60QD2_9b5kP5W4Mt5W21N4q9tW7Km6kG2pjpZPW9f0-GR6_C3RLVYd6LC7LGhRvW7YmfCp1_xcTSW4JNMr33bsLpjW7mdglL569p2xMssSLKZ_Wc7W65ljLW3Wrl1hW84XCly7hSxHZW87CF0z4lwQxMW6v2Nw-7xR94tW6DGHVQ2jn857W7FPV_q76RMzsW99mxs_4fGXMvW8cg52j8d1hDpW5MvrJn5RcXcGW83glV05l-5mRW4r2KTj1xCWnNW49Bzmy7XqtMWW2Y3qF47bRPXDW6nHx592MpnKpW73y5NL1LcMz7W49HGfZ28RJFtW61Cpqp5C0Rjqf1CKkLY04> from 2022 to 2025. It showed that 30 percent of uses were related to work while 70 percent were related to non-work activities. In addition, the gender gap among users among users shrank steadily during that period. - In January 2025, Anthropic’s study of Claude <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3n2W3FZwH72gQLPmMDFTCMKTk4JW8Sp0Dm8m1h3NW8RH7W_3V1lL5W4lHG7X5MX36_W23hmkW5t954HN1DfYFhV0TkBVh4gP329XMV9W82F4JW3c8qCmW1wfyqy4z-kVtW44zxzn8cl1ZhW4by7Yh8_szdzW4hP3M07VxSjHW4nQczv18c48BW1ss5yQ59yckrW760lpp6WW-5lW2lhgnj292JLdW6zZ1yZ7Bhg3WMJRhrdCk4LdW7dzMVK1Q-mkKW32N-Mb3SXM5fW43YmQ560pSdtW8x1HFj8L6k4xW2dSKHF4pBNSDMGr_3q7z1ZyW5GPGcj98lTFNW1vDbcF9gjmwdW735N-b2VC6QtW4hk2sF2XNm9SVT9Z235BP9zCW3RPKBt1xs5zsW4DhHvy5JMT3Ff4FzF6404> showed that the model’s user base focused on work, especially software development and text communications. A small but growing number of users engaged in games like Dungeons & Dragons and sexual roleplay (despite prohibition of that use by Claude’s terms of service). Why it matters: The authors argue that the AI community may need to rethink chatbot design altogether. If users treat chatbots differently on mobile and desktop devices, AI builders would do well to design their systems to suit the devices that will deliver them. Application design is one way to accomplish this, but system prompts may be another. Desktop chatbots and agents can respond with more information-dense answers, guiding users to execute tasks, while mobile agents can offer shorter, more empathetic responses. We’re thinking: Studies of chatbot usage conducted by different companies show different results. Perhaps each company’s users treat AI differently, so the results of any given study may not apply generally. That said, the Microsoft study suggests that the device used and the time when it’s used can have a big impact on what users want — important considerations for designing any application. [image: Diagrams comparing LongCoT and Delethink environments show reasoning processes and context management.] More Affordable Reasoning One way to improve a reasoning model’s performance is to let it produce a longer chain of thought. However, attending to ever-longer contexts can become expensive, and making that attention more efficient requires changes to a model’s architecture. Researchers proposed a way to limit the cost of processing long chains of thought with just a bit of training. What’s new: Delethink <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nKW8SXyhJ7DQcqSW8nHYDW27ZnkMW7-qqxL1DZ4LXW99kBsN2HH6cXN3hSL0PzN9_GW88SD2t62Xn-NW7jC2pZ5GwK4CW5NRxGv8YSxJpW58GnTd4-W-PbW4VG01n68vhq3W3c4Hvz3trwNJN5cdb0q5KsgxW6q-32N7fK-0lV4JBTF5qrtGGW9jgDm36dLyBMW91y9R58pXr96W69J70p7dz_LvVbq9_b5gbqjmW12FCCy3YdKTPMWv2gLQgnBKW2Qvm1-1VWPlsW8731bg6JnDL0N3Ry2sX9pj82W2l-hmn1M1NwDf8Qx6z404> is a reinforcement learning (RL) method that trains large language models to periodically truncate reasoning tokens to a fixed maximum number. The authors include Milad Aghajohari, Kamran Chitsaz, Amirhossein Kazemnejad, and colleagues at Mila, Microsoft, McGill University, ServiceNow Research, Polytechnique Montréal, and Université de Montréal. Key insight: Reasoning tokens typically accumulate within a large language model’s context window, where they consume quadratically more computation as the contents of the window expand. One way to counter this effect is to train the model to reason within a maximum context window size. In effect, as a model is reasoning, it can learn to replace its chain of thought periodically with its latest “thoughts” and then continue. How it works: The authors fine-tuned R1-Distill 1.5B, a large language model, on math problems in the DeepScaleR dataset <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHq25nR3bW69t95C6lZ3mTW80Q7xZ4QXs__N1lxhWpcckqyW8wVFFy6hzJcVW7lSDK17bJLCDW4lhZLT5Sd68CW9h4NJR60THQ-W5k4NRL8Q42SYW80l_MJ80xtntW6wHP2T7R1_QvM3MrcZJQBj4W7sKMf89hjZJZN7mGhVXYlC5JW75Xhct8CB4X6W48SvnV7pRGdlW3LDlbC5mZnl1W1lM38N4LctGlW4Kvmxd7_MBlFW2D9djG45RMj6V3PgnB349h89VDJ0N86CN-X-W91qhB675mcPDW3Xv7FR3zltJCN6mtqCgyZ0QYW1h0rpD6KVCZYW2lzJK27Fp93jW3rtB8P6p3RQbVT5mqm8LrDz8W4-kgsq6p-6fTW3MytCj3R3WLpW38hspr2T_sqRVXXhZ_3fxpH4W1s9Xp_36n6YCW9681k46Q2Cn-W804H7f3-t3N3W4ld37c4JpVXDW6dp7BY8lg1Z5f3ztR6-04>. They used a modified version of the reinforcement learning algorithm GRPO that trained the model to reason in 4,000-token chunks: - Given a math problem, the model generated a chain of thought until it had either finished or filled the model’s context window with 8,000 tokens. - If it didn’t finish its chain of thought, the authors replaced the context with the original query plus the last 4,000 tokens. Then the model continued to generate its chain of thought until it had either finished or the context window once again held 8,000 tokens. - They repeated this process until the model had either finished its chain of thought or produced 24,000 reasoning tokens. - Then the model attempted to solve the problem, receiving a reward for a correct solution. Results: The authors compared their R1-Distill 1.5B models to the same model after fine-tuning on the same 24,000-token reasoning budget via using GRPO. They tested the models on reasoning budgets of 24,000, 96,000, and 128,000 tokens. - With a budget of 24,000 tokens, their model matched or surpassed the baseline on all 3 math benchmarks tested. For example, on AIME 2025, Delethink (31 percent accuracy) outperformed the baseline (29 percent accuracy). - Their model’s performance continued to improve as the authors increased the reasoning budget, while the baseline achieved much smaller gains. For instance, with a budget of 128,000 tokens, their model achieved 35 percent accuracy, while the baseline achieved 30 percent accuracy. - The authors estimated that training their model with a 96,000-token reasoning budget would cost 7 H100-months, while the baseline would require 27 H100-months. Why it matters: This work eases the quadratic compute barrier that can make extremely long reasoning computationally infeasible. While other methods, like linear attention, achieve the same result by changing the attention mechanism, Delethink restructures the reasoning process to limit processing regardless of a model’s attention mechanism. It opens a path to reason efficiently over longer contexts without requiring new model architectures. We’re thinking: As the authors mention, most LLMs are pretrained using relatively short contexts. For example, Llama 3 models started pretraining with examples of 8,000 tokens. This may have made them good at processing inputs around 8,000 tokens long. That is to say, Delethink’s performance may have been helped by the fact that LLMs tend to be pretrained on short-context tasks. Work With Andrew Ng Join the teams that are bringing AI to the world! Check out job openings at DeepLearning.AI <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3p5N65swpFxLdcBW8s0KY56mqGLPW54Gqb655n5yLW2VlKw93KfzxyVsN0Xb6JKmjBW58N5lm172PBQW1pyv7s2nLPmJW425jTD7Kg0ltW6L4phx2-p_-fW4f56hZ8KXv_QW92xSwf4_ytncW4592Ww3PryZ1W8573kw11T8gtW11NgFR1dbB96W3YbCdl4xQrpNN81SyzgkdP_TW3X5ZmK4ZtpbnW8SFYXH5z_h9FW56-zy35zYthhW7Z8hQ14HdSpGW86N-fM4Fr3J2W6sVzYm75ZmQDW1yYb9B41ckycVXn2j82_vpHcf6T8cZl04> , AI Fund <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lbW49jm1g3GPY-6N6cnvR7rjkr-W8VgP9N5JN8qsW6yShD98ZcZqKVrKLqv3dPMCpW718j8y5JdRKMW2tsL9K4nVRPnW686Cl16p2Bc4V9V3rl94vpRvVm48ql1qZpmxW68CxMZ5CfhzDV5KS6l8P2GyXVWcLT93KYJZcW4cs8CQ4dWkV_W8-rzXF4LCJjWV7Ltx08PgvF6W3ZKBYn7LjkykW4w4bvB4b9dgpW5q3QZW987twRF903KzsnMM2W7r5RZ12C1vQWW6bZJhw9htKGMW8lFfhm2DbYx_W7-NxNX7GdCfMf3Z3kkT04>, and Landing AI <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nTW1lpkY37j3bCYW33RpBy4V29pwW98gl3t4BfKvGW4zYqFT5Qjxl6N1RXgGSVm1WrW7xP5T35MzQ_LW8rWqPX5ZW-dKW4DVSpQ1S4YbXW392rhC2-v-X3W6Fs7xg7RtQVdW9llkVF7-jWgmW2Py9Qd71P794W77Js7c6wGQy9W1mDLHH4MFp59W7FbMRR989XsQW97sLMj6YmLRbW3GXV_R3XnjFyW2qLwG01d0fwXVbzPfm2hWP9hW2Y8Htq6wMdHkW8PqLSG2S1-NjW8G4rfc5dtp1SW8jKk_Z44jcmMW6WBJ4K1Z0S1_dZvf6804> . Subscribe and view previous issues here <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3prW7Qb8g-2wx0YNVy8LrQ4CN6dbW3MHBx77698fhW8QshbW4N0vzqN4Prl3f180wZW1MF8S644trRSW1SwZL88hTcWfW2t4t953T985BW7vH8P03bfd35W86y-3F1K69nqW8Tg_6s4rlkY0W5_7hHx7fv5-RW1rtZ032W9C_xW3-NFd_79zfRNW1WgY_z5R8lQzW7YzSr-2J5t3TW59mPLM72YyqMW2XhG3v2PBjKcN2Bj7GGwPWv6MYndxzMw1qGW1N2Vmb1Qr12jW51_Yxt7LG-W1W1tjyPX7mh_KhW7LJPvS60H8jDf1gjZhP04> . Thoughts, suggestions, feedback? Please send to thebatch@deeplearning.ai <thebatch@deeplearning.ai?subject=RE:%20Thoughts,%20Suggestions,%20feedback>. Avoid our newsletter ending up in your spam folder by adding our email address to your contacts list. DeepLearning.AI, 400 Castro St., Suite 600, Mountain View, CA 94041, United States Unsubscribe <https://info.deeplearning.ai/hs/preferences-center/en/direct?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159> Manage preferences <https://info.deeplearning.ai/hs/preferences-center/en/page?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159> -- Markus Edgar Hormeß Founding Partner WorkPlayExperience GbR Oedenberger Weg 14 D90571 Schwaig b. Nürnberg Germany +49 151 24101032 markus@workplayexperience.com www.workplayexperience.com
HTML
<div dir="ltr"><br><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">---------- Forwarded message ---------<br>From: <strong class="gmail_sendername" dir="auto">The Batch @ DeepLearning.AI</strong> <span dir="auto"><<a href="mailto:thebatch@deeplearning.ai">thebatch@deeplearning.ai</a>></span><br>Date: Fri, 9 Jan 2026 at 12:03<br>Subject: LLMs Go To Confession, Automated Scientific Research, What Copilot Users Want, Reasoning For Less<br>To: <<a href="mailto:markus@workplayexperience.com">markus@workplayexperience.com</a>><br></div><br><br><div class="msg-6958527957600471347"><u></u> <div id="m_-6958527957600471347hs_body" bgcolor="#ffffff" style="margin:0!important;padding:0!important;font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word"> <div id="m_-6958527957600471347preview_text" style="display:none;font-size:1px;color:#ffffff;line-height:1px;max-height:0px;max-width:0px;opacity:0;overflow:hidden" lang="en">We just launched a course that shows people who have never coded before, in less than 30 minutes, how to describe an idea for an app and build it using AI.</div> <div lang="en" style="background-color:#ffffff" bgcolor="#ffffff"> <table role="presentation" cellpadding="0" cellspacing="0" style="margin:0;padding:0;width:100%!important;min-width:320px!important;height:100%!important" width="100%" height="100%"> <tbody><tr> <td class="m_-6958527957600471347hse-body-wrapper-td" valign="top" style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding-top:20px"> <div id="m_-6958527957600471347hs_cos_wrapper_main" style="color:inherit;font-size:inherit;line-height:inherit"> <div id="m_-6958527957600471347section_1706730515171" class="m_-6958527957600471347hse-section" style="padding-left:10px;padding-right:10px"> <div class="m_-6958527957600471347hse-column-container" style="min-width:280px;max-width:600px;margin:0 auto"> <div id="m_-6958527957600471347column_1706730515171_0" class="m_-6958527957600471347hse-column m_-6958527957600471347hse-size-12"> <div id="m_-6958527957600471347hs_cos_wrapper_module_17067305151702" style="color:inherit;font-size:inherit;line-height:inherit"><div style="overflow:hidden"> <table role="presentation" width="100%" cellpadding="0" cellspacing="0" style="text-align:right;font-family:Arial,sans-serif;font-size:12px;line-height:135%;color:#23496d;margin-bottom:0;padding:0" align="right"> <tbody> <tr> <td align="right" valign="top" style="color:#3b3b3b;word-break:break-word;text-align:right;font-family:Arial,sans-serif;font-size:12px;padding:10px 0;margin-bottom:0;line-height:135%"> <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx5nR3bW95jVnq6lZ3m9W4FHPhr8rhpmXN4bY_GGr2cxDW3JmQfL6MvqJ1W8N-Slr32QYRCW5tw1Fs8S11NRW6YwScS2ls0H0W8VmJJ92V6zZLN5kxtb_WssNwW22YZlN1vpFWgN2RYcnvfd9ScW1f6jc-5Cqm33W46NtVN59ZS0XW81-zG45dNvyDN7rmc1QGRg92W2nD_Bh47DC3yVxtVWs5JRV-JW965Pvs7Sp7WwW7lhHNF21ns0ZW4h2PSP94cCFWW3rNvJK8z0kf7VXQ95L55R3pgW3DC3tB5wR5mgW8GcYp18sMf0yW7PFNHD93ndzYW21KLx76Ft_txN77pF4pjPwtsN8k8yk1XtVc1W4rR-dX2XRN3nW76gTJ98sCM1CW6Hm1r-2nMFVhW1lxxdk8Zgts3W7D3ZFv9m0tFyW6Zpw0k7pKDsbW34ZSVn2nwNPFM3791TH8ZJXW8gTzff43q23ZW3dvKsz5TNcFlW5NYthl7qPq4pN3J0q-CrPjP1W44rHQ854vR2PW3-n0FD1Xcb6_V41BBx7Wz9L7W7mq1GH6Vz6NYW954P6513wyynW8nBMsK3pgtvMW3Bg8pQ6vsdgnf8B1frb04" style="color:#237b94;font-weight:normal;text-decoration:underline;font-style:normal" target="_blank">View in browser</a> </td> </tr> </tbody> </table> </div></div> </div> </div> </div> <div id="m_-6958527957600471347section_0" class="m_-6958527957600471347hse-section" style="padding-left:10px;padding-right:10px;background-color:#ffffff" bgcolor="#ffffff"> <div class="m_-6958527957600471347hse-column-container" style="min-width:280px;max-width:600px;margin:0 auto;padding-bottom:10px;padding-top:10px"> <div id="m_-6958527957600471347column_0_0" class="m_-6958527957600471347hse-column m_-6958527957600471347hse-size-12"> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:0px;font-size:0px"> <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lvN4B5SvcbldLVW2KNgKQ4ZjL91W8HSh4212Cxb2W7yZy8K6JxxJYW3QJ4dw1-YPkxW6BPHMY2DgQ7PW4FPL221rbj_BW47BBdg6CqfcsW3yLgwB7r_K1HW6NXWtz2Dh6k5W99fcQp5YB-RcVnL-bJ3QcYWJW6FZTMB7NBhjhW1qhCfF6D6dy1V1nWkW5RJNtPW3jqW3G3Q8L-WVdq4217X2dljW3KnMvV98BR69M_KgHYgls0MW8yJKmj6_4yHDN3tJ2kt4q_9dW2GkMLj53rrz4W73zM0W4yNFKgW67NFZZ85lNBvf59sbLT04" style="color:#00a4bd" target="_blank"> <img alt="The Batch top banner - January 9, 2026" src="https://info.deeplearning.ai/hs-fs/hubfs/The%20Batch%20top%20banners%20(70).png?width=1200&upscale=true&name=The%20Batch%20top%20banners%20(70).png" style="outline:none;text-decoration:none;border:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </a> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old1" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old1_" style="color:inherit;font-size:inherit;line-height:inherit"><div style="line-height:125%;text-align:center" align="center"><span style="font-size:15px;font-family:Helvetica,sans-serif"><a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3mDW128n0H27S-49VngKyZ59WlPCW8rvm_06Js47MN26yq4ltTSgnW22Syvb4-FzZCW4yHs6L6ZZp4YMb4Wv4whLQsN20TjVNWnTNnW2pJBlK4NB0vKW84jDhw1jgxRXW4Wmr875hRCVzW2RdM7086dXRjW53GdyZ297BNDW7n1SZp60p-mJW4m_89W79tr8PW2vRF0G2wGTR1W4H8jL56R9vnGW5zBc6s8X6cZlW8rLTm07q3RDlW29SNyl6Hq3n5W83KlcS65zxrRW5kcrXd2rctX4N4SHqZGctbw6W7brsFB96DmjMf7hN02R04" style="font-size:15px;color:#f53b0d;text-decoration:underline" rel="noopener" target="_blank">Subscribe</a> <a href="mailto:thebatch@deeplearning.ai?subject=RE%3A%20Tips%20and%20News" style="font-size:15px;color:#f53b0d;text-decoration:underline" rel="noopener" target="_blank">Submit a tip</a></span></div> <p style="line-height:150%"> </p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-family:Georgia,serif">Dear friends,</span></p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-family:Georgia,serif">We just launched a course that shows people who have never coded before, in less than 30 minutes, how to describe an idea for an app and build it using AI. It is now time for everyone — marketers, product professionals, operations specialists, analysts, students — to build software applications with AI!</span></p> <p style="line-height:150%"><span style="font-family:Georgia,serif"><br>I’ve often spoken about why <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3n-W1mny-F56FdD0W1hcc0c7mP1dYW2TSZhj83lZK_W4p2y2J7hN1bPW3cShVp3pZ2rvN2K178TBd__BN6bKqT--9GL4W66HXht7VTMndW5kFyTL18ZSY7W6dNSqv4zhnxKW8mp-vV2bw2LvW5BHnk-3M_7GgW6KycxS677nwzW3_bRw-5Gpj5PN8wGfbFz_yFNW30ZZpw26vq1qV-xStt31rggtW2rxTHZ5TFrXGW60mvHB8k2ZDSW7HyXNM1WY5V0W4rKwhs8cmZF9W2s0pHq7ckmF0V7RW4c8hlN9DW6CYjqJ7-d-45W4SMhcK6dN3dcW4-RZ2C82dmv1W1WDmS12H8zM_W4xF-1Q43SRRVf8vfV3P04" rel="noopener" style="color:#237b94" target="_blank">everyone should learn to code</a>. I’m seeing a rapidly growing productivity gap between people who know how to code and those who don’t. For many job roles I hire for, I now require at least basic coding knowledge. Many times, after I speak with a non-technical audience about the importance of building software using AI, people ask me how to get started. In the past, I didn’t have a great answer. </span><span style="font-family:Georgia,serif">T</span><span style="font-family:Georgia,serif">hat motivated the DeepLearning.AI team to create “</span><a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3ndW1NFm7J4cS7HKW5-lT9S5G9KWQW3Kl8q36Lnm9qW8xBqQ91d9ltcW3c_ztd5Nj6fbW8_-wBW5lqVKxVJjGWn8gw6bjW50b5FY2SXBYrW6Rv3hn1hVDXbW9148Jv385H8XW1bDpHK8Dt0cJW2TGJF295ZhLmW5brfWH5JFWmMW100VHl1nR409W7njgC79jmb9GN2rkT_hQ9Xx0W3wZ87N3KgSP5W95mH222mxX_CW1PxCg475GFVQW2fqFFM236wR8W2Vm0M_7kzgyYW4R7Jtd6hFYdJW4gmbsW22Z7q-W4jf5r12-j1-SW7x1Spn3ht80GN3G989xnnzZ6f4-PBsH04" style="color:#237b94;font-family:Georgia,serif" rel="noopener" target="_blank">Build with Andrew</a><span style="font-family:Georgia,serif">.” It’s the best way for someone who wants to try vibe cod</span><span style="font-family:Georgia,serif">ing to get started!</span></p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-family:Georgia,serif">This course requires no prior knowledge of AI or coding. And it’s vendor-agnostic. Specifically, learners can use these techniques with whatever tool they’re most comfortable with (like ChatGPT, Gemini, Claude, or the chatbot built into the DeepLearning.AI platform).</span></p></div></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old2" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:10px 0px;font-size:0px"> <img alt="A birthday card generator form shows fields filled with humorous data and a chat bubble indicating help needed." src="https://info.deeplearning.ai/hs-fs/hubfs/2026.01.09%20LETTER.png?width=1200&upscale=true&name=2026.01.09%20LETTER.png" style="outline:none;text-decoration:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old3" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old3_" style="color:inherit;font-size:inherit;line-height:inherit"><p style="line-height:150%"><span style="font-family:Georgia,serif">If you take this course, you will build a working web application: a funny interactive birthday message generator that runs in your browser and can be shared with friends. You’ll customize it by telling AI how you want it changed, and tweak it until it works the way you want. By the end, you’ll have a repeatable process you can apply to build a wide variety of applications. </span></p> <p style="line-height:150%"><span style="font-family:Georgia,serif"><br>DeepLearning.AI’s mission is to empower everyone to build with AI. This course is just one of many steps in service of this mission. <br>If you are already a developer, please encourage your non-developer friends to try their hand at getting AI to code for them. Not only will this help their productivity, they will find it really fun as well. Please invite your friends to come <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lxW6Z-p6p842JrLW7yg_Fx7y0RthW4tZ2YR8bkSbFW2VWgNL3ppRTxW3pyCzV1rN1mJW3pDnwr4L9PHYW8TkLch3GsmnWW2Sv6rs87XlvjW5YMPmc5RG9KvW6LKz0C5t_D88W3CP8xy7d8LC1W7YsdXZ1cSFnnW44yGMF8g0xYBW9kqBQK4P9MjcW6G3sdy7KLffjW8VcD9_2lyJXHW5y_H-b6Y8N0fW6jWkhj1RykwWW8nt0-w6rqMNTW3944c41G6PG6W68_j_y3yL8bWW41llRh51L8lsW7ctSv-26xvF6VhRWQl3HPRLcVr9jjb3yjl4_W7crrh18zqYQSf6HXPMj04" rel="noopener" style="color:#237b94" target="_blank">build with me</a>!</span></p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-family:Georgia,serif">Keep building,</span></p> <p style="line-height:150%"><span style="font-family:Georgia,serif">Andrew </span></p> <p style="line-height:150%"> </p> <p style="font-size:18px;line-height:1.5"> </p></div></div> <div id="m_-6958527957600471347hs_cos_wrapper_module_16890053646291" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_module_16890053646291_" style="color:inherit;font-size:inherit;line-height:inherit"><h2 style="margin:0;font-size:21px;line-height:150%"><span style="color:#000000">A MESSAGE FROM</span> <span style="color:#000000"><a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql3qgz0W6N1vHY6lZ3pNN432hYGKZ9k9W65ntXq84NPj7W6QdjW-6jjJKTVSfVsG5bKFyLW5k_4LL1C0qc3W3jG5w38645YMW3Vwvdz5CMtqvW4wxhq82mXJtvMd26xqzx1jhW6C2qkT5CQH1YW7Kl5Zq8kkB3CW9b79PR7WLD_nW3_Z_Tv6-nJY9W6Ybjsj6qqdC5W7Tp1XG1WF_HFW2071jr6JqsxdW7JW4hH1HYc5mW5RZh736dJGS7W17sHWC6zkhK5N8klj4jtTH5PW5YJ0Yw4MQL-XN4FZn5dSyYQzf4XFBW404" style="font-weight:bold;text-decoration:none;color:#000000" rel="noopener" target="_blank">DEEPLEARNING.AI</a></span></h2></div></div> <div id="m_-6958527957600471347hs_cos_wrapper_module_16848647586863" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px"> <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3n_W2n5v537npkmlW34qHlf41HL6SW77FH7H6W9QttW8f8JCk6txg6HW5lww2h30YTYYMKnHkKTVbGjW3YFdWh1NFVc8W4sm7Mh9d4P3GW8cstBj6rGbTYN2gcvBhKGScpW7T8Vck8zsQqYN15mkhN19Y-_W1zbxz27-zjJtW7cvD9G57r7K7W5K_7pg5pt7XxW6LPKx_64ngJxW7yd5Fg1y1q5kW62qlvj5s1q0-W6lxjVL5RRS8BW7v6mX16g2Bh6W49w4tk2y1d-wVD8qp77KxYtcW7ChZSC8vk9FnW8ySxlv2Hm2bTVV4FT669k2xqW9jVRbS3XSdvTf4j1sxC04" style="color:#00a4bd" target="_blank"> <img alt="Promo banner for: "Build with Andrew"" src="https://info.deeplearning.ai/hs-fs/hubfs/Build%20with%20Andrew_Banner_2_1920x1080-01.png?width=1200&upscale=true&name=Build%20with%20Andrew_Banner_2_1920x1080-01.png" style="outline:none;text-decoration:none;border:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </a> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old19" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old19_" style="color:inherit;font-size:inherit;line-height:inherit"><p style="line-height:150%"><span style="font-family:Helvetica,Arial,sans-serif"><span style="color:inherit;font-size:inherit">You don’t need to learn how to code to build an app. In “<span style="font-style:normal">Build with Andrew,”</span> Andrew Ng shows how to turn ideas you describe in natural language into working web apps. Perfect for beginners, and easy to share with someone who has been waiting to start. <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3pmW8QHvqM4KhKGnW8T2m_K5z0fxnW6TsRVR2sJhynVX1TBS3RXnDhN4pD_vWBZ0ZtN1Zw0FJb7pQ-W6h2Vcc5mZp4XW550B0L8vlgd4V7F9G-2MzZ4fW8v-mZL7ZK5RLM83LrPP9wfkW6r6LWP1CBx9bW7n2Csx8FDDllW6chw1v64kfT5W3RPsXW8D8ytkN2wgWCy4kFpBW2dXNXg8B3m8fW9lFd_x2fmRG9W8fKcPb8j5d6JW48hDw56dxDGhW7FcD4Y5-MvCRW39lSTL66fcVNW9dkc8N329Yb5W2fXT0P5x06QhN3gyCFHPnlV1W1SFhf72rdZvff4r4PWM04" rel="noopener" style="color:#237b94" target="_blank">Explore the course now!</a></span></span></p> <p style="line-height:150%"> </p></div></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old7" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old7_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;font-size:32px;line-height:125%">News</h1></div></div> <div id="m_-6958527957600471347hs_cos_wrapper_module_16195246841011" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px"> <img alt="Dialogue displays a model revealing it answered incorrectly and wrote code against instructions." src="https://info.deeplearning.ai/hs-fs/hubfs/CONFESSIONS.png?width=1200&upscale=true&name=CONFESSIONS.png" style="outline:none;text-decoration:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old12" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old12_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold">Teaching Models to Tell the Truth</h1> <p style="line-height:150%"> </p> <p style="line-height:150%">Large language models occasionally conceal their failures to comply with constraints they’ve been trained or prompted to observe. Researchers trained an LLM to admit when it disobeyed.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">What’s new:</span> Manas Joglekar, Jeremy Chen, Gabriel Wu and colleagues at OpenAI fine-tuned GPT-5 Thinking to <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3pZW7gKR221JKmCxVZ_gw13WSktRW2CwXWr8_dhr1W6yB4qg8VRbS5W2Ljd8m5RN0LkVJy-9y2NxCVTN539nfWKBvHCVwkKqN8636n7W4jThy671t1BNW3RPK5s3z9WbkW5n2JPd7sNF_cW5XcmT03ZMzYqW7qRklg37s8tWW8_-pz58N3mqgW5Ww9tg42dG5WW8fF8xm2mp2ZZW4LQC1Y3VTnX7W65Z3ZC4NqL_XM3NxWdVDkW3W8X8WmJ6K0KmkVKD83f5t1B6-W5MM08Y5gTghTN3nFLgN9SptNW5GDYrF8wH4y8N3DDM8M_LTp7VL34KM7DrSx3W3cwjKw4r1P6dW3R4hjL44l8wLW6jHrNd4dbvsNW8RnV_W4B-QS4f4HmDBb04" rel="noopener" style="color:#237b94" target="_blank">confess</a> to violations of its instructions or policies. Upon request, the fine-tuned model revealed any transgressions it had committed while producing its most recent response.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">Key insight:</span> During training via reinforcement learning to encourage accuracy, style, compliance, and the like, a model may inadvertently learn to prioritize earning rewards over pretrained or prompted objectives. Under some circumstances, this may cause it to hallucinate, lie, scheme, and otherwise misbehave. Rewarding the model for confessing to its misbehaviors in response to a designated prompt should train it to admit its mistakes without affecting its performance.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">How it works:</span> The authors used reinforcement learning to fine-tune GPT‑5 Thinking as they normally would, and also to produce confessions.</p> <ul style="line-height:150%"> <li aria-level="1">An unspecified percentage of the time, given a prompt, the model was rewarded based on its response’s “correctness, style, helpfulness, compliance, safety, and more.”</li> <li aria-level="1">The rest of the time, after the model had produced a response, the authors prompted it to produce a confession. The confession described (i) all constraints, instructions, and objectives, implicit or explicit, the response should satisfy, (ii) how well the response satisfied these criteria, and (iii) any ambiguities or uncertainties in the criteria.</li> <li aria-level="1">Given the query, response, and confession, an unspecified reward model rewarded the model based on the confession’s completeness and accuracy — a measure of the model’s honesty in confessing, since items that were missing from or falsified in the report can be thought of as lies.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Results:</span> The authors performed 12 evaluations of the model’s ability to follow instructions, resist jailbreaks, produce accurate output, and complete an impossible task (coding a script that required essentially 0 seconds to run). The fine-tuned model, when it misbehaved, confessed to its misbehavior at least half the time in 11 of them. For example, across 5 evaluations for hallucination, it either didn’t hallucinate or owned up to its hallucination 81.4 percent of the time:</p> <ul style="line-height:150%"> <li aria-level="1">The model did not hallucinate and did not confess 42.2 percent of the time.</li> <li aria-level="1">It produced a hallucination and admitted the mistake 39.2 percent of the time.</li> <li aria-level="1">The model did not hallucinate but falsely admitted having done so 11 percent of the time.</li> <li aria-level="1">It hallucinated but failed to confess to it 2.9 percent of the time.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Yes, but:</span> As the authors point out, even when models are forthcoming, confessions are not a panacea. Models won’t confess to behaviors they don’t know are wrong.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">Why it matters:</span> At inference, confessions can be used to monitor a model’s actions and stop undesired behaviors. <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3l9W5KRCkR5T024KW6nYrjD52VPH2W3PRwrr5Jsm_KW235qMP1M3mKyN5WZtf219N-9V5bpp28lTMwdW9lVRF23jxWRzW897JTY1fBS4_W21-GQX4bgQzZVC75Rk7q2KB1W245qX23CS7hVW6CNwC532wRS7N4Trj6tmHpCJW8xr_8R8nc_X9W7DrrHY3WVPb3MmfWlxzmDlgW8hGT5w8wp1PzW1dmlLy7vkLpCW548h5L3VJx6JW7XHPS86Z87ZSW87t6xB3N682KW4GD5nm7KNC1nW7vT5dL9fRWLrW2hVPry1KnLPVf1Wn4Wj04" rel="noopener" style="color:#237b94" target="_blank">Chain-of-thought monitoring</a>, which classifies bad behaviors a model might describe in its chain of thought, can be used the same way but, unlike that method, the authors’ approach trains models to reveal misbehaviors they may <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lvN3_CC9wCqgxDW7ZDn5b7PMBMjW53WXJc3nFJL2W4n8-TK82qTQBW1Rlqxl50YJFZW2Wp-CY9ghbHmW7XLDF41CBx-wW4-6w_C8RdQ_DW8RC9LX7_vtgBVZplLq176XHSF8XGqRsWKkNW6gmF9K1Zp-P7W8-1gnX4_-8pPW7RqF0T5HQ5wwW98-vyK1lGZGSW35hNJL5CtfSYW47Fnk_8yCKMNW1tN0zc53N-8pN1Mbl3JqMsQRW1p8Tdf2tQ4VWW9hG0HB1Q80y-MXMqvpZMcWhW28tVk56jh6xdN8qQ92FGBM6wW13q9S0814y0gW4Sr_KT44SD9XW6z6w_t5tWWkgW4DFYsq8xGGt0W1qQVns1tD71YW4GYw-W5GRCQkW493scf5bBPr-Vjfm102ZbtkfN6L93nSNj46kW5vn7913yKm2Xf7Y5K4s04" rel="noopener" style="color:#237b94" target="_blank">omit</a> from their chains of thought.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">We’re thinking:</span> We always hesitate to anthropomorphize model behavior, but this work may be a step on the path to giving AI models something that resembles a conscience.</p> <p style="line-height:150%"> </p></div></div> <table role="presentation" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px"><div id="m_-6958527957600471347hs_cos_wrapper_module_17370396178185" style="color:inherit;font-size:inherit;line-height:inherit"><table role="none" border="0" cellpadding="0" cellspacing="0" style="vertical-align:top" width="100%"> <tbody><tr> <td align="center" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word"> <p style="font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto"></p> </td> </tr> </tbody></table></div></td></tr></tbody></table> <div id="m_-6958527957600471347hs_cos_wrapper_module_17241662452411" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px"> <img alt="Diagram showing SCP hub linking clients with databases, tools, AI agents, and lab devices for experiments." src="https://info.deeplearning.ai/hs-fs/hubfs/SCP%202.png?width=1200&upscale=true&name=SCP%202.png" style="outline:none;text-decoration:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old9" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old9_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold">Lingua Franca for Science Labs</h1> <p style="line-height:150%"> </p> <p style="line-height:150%">An open protocol aims to enable AI agents to conduct scientific research autonomously across disciplinary and institutional boundaries.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">What’s new:</span> Shanghai Artificial Intelligence Laboratory (SAIL) published <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3m9W1RnzSC6ptRNnW8lQhT48hv5k-N5Zs97xCjX4lW9g_Dhm6DwcL5W8vBWxx7t0_HSW8M8p6q3-dnQLW5J0XzX398YzqVf5Bc54bx8yDW524wQQ58zlssW8C-VsF4JPk28W2X4J8N1qf0LyW33Vnyx49JqxrW6NL48t1b9rj8W26KFC78Hfw8-W7fkR7S1RXWnFW93n0St3jqkxXW3C6-DT1N-YcWW79hk8631cLlNW8QDR9c7nQfrlW7Q33XB5H44FqW73Sjlt5ZscYqT50hM99rzwvW4jR8Wx7nf980W7pTF5-62Tc6gf233G-H04" rel="noopener" style="color:#237b94" target="_blank">Science Context Protocol (SCP)</a>, an open-source standard that connects agents with local clients, central hubs, and edge servers to conduct automated scientific inquiry. SCP is published under the Apache 2.0 license, allowing commercial use and modifications.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">How it works:</span> SCP attempts to make experiments using AI agents and robotic equipment as reproducible as possible. Like <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p_W80pf4-6yt73mW8-RbY95mPKsbW87Sl6s8H32BgW8b2ZKl1yKJyHW6knHVv8wptQ_W6NJfTv6HB9J9W7PHwFz4t6k71VCScHP4PZm8pVGd47X3hNzDlN87slVCsP08mW5rc2gd5XKYgVVmTkkS3TlS0mW4b6j6Q99JG7KVzVpVn3ZvztkW4PY9SB4Jjp0hW4PrwQp8sPFrLW4M0cR-6spBrvW3WjWl75Ym2-KW131g4y85HlqLW7YGRfN3d9V5VW6n2BqC7Gbjc4W7Y1xlJ2qNtJbW7Yw_cH3Y-s4kVFwqcz1f9t0MW2N9-ch5hr_1RW7KzxMq6phDk_W660yn_3Lh1twN9cpQwrqhsWfW5fgcRK5tXZ6KW75Mf8t37XBfLW8l2FNB3fM6HWVm8Wtw5pkqZqf6N23J604" rel="noopener" style="color:#237b94" target="_blank">Model Context Protocol</a> (MCP), it enables agents to interact with external resources. Unlike MCP, in which servers stand alone, SCP’s design requires centralized hubs that manage other servers as well as the client applications that enable users to access them. In addition, SCP’s structure offers greater security by governing messages and tools more strictly than MCP, which is necessary in scientific experimentation, the authors say.</p> <ul style="line-height:150%"> <li aria-level="1">SCP’s fundamental data unit is an experiment. Every experiment is stored as a JSON structured data file with a persistent identifier and record of an experiment’s type, goals, data, and configuration. The format makes experiments traceable, versionable, machine-readable, and consistent with institutional policies that govern data.</li> <li aria-level="1">An SCP client authenticates users and gives them access to institutional resources. Researchers can describe an experiment’s goal in natural language (for example, “increase the brightness of this fluorescent protein”) or upload a complete research plan in text or PDF for their hub to analyze.</li> <li aria-level="1">An SCP hub takes a goal or other request and uses large language models to generate a set of experimental plans that list steps to carry out the experiment. The hub measures and ranks each plan according to its resource requirements, cost, and risk at each step. The user selects one plan, and the hub then orchestrates and schedules multiple agents and servers, which carry out the experiment. After an experiment is completed, the hub archives it for researchers to consult, alter, or repeat.</li> <li aria-level="1">Edge servers manage the experiments planned by the hub and stream data back to it (which in turn returns data to the client). Servers may belong to an institution, or they may be devoted to a particular discipline like biochemistry or mathematics, each with its own specialized tools and databases.</li> <li aria-level="1">The protocol currently includes more than <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nsW8S6nn78LHn7yW46FC401xlyVBW11xmKC2kTLbqW8BnVND2X40nLW1pfSh14-_05dW1xTQgS5NnbrNW3Br4Gw2tpp9VW2KXNjq259l14VWdW3_49PFY9W37ZcFx62f96PW5wlMNt7rG3vZW73J8v11Vrf_9W5QKBCp12qgSSW7WyjZ-40RbzcW5dVYLh6QpWD4W4CqYjw4LP1-gW57FwRm7-1NBsW1nLBL87Pln86W2_SCkz9bXDWPW7ZBWKb4gK8d2W44Srxz5w91mzW1SGZlH64qQv-W6PCm_k8LcZ3tW7s9m8-6dRP8yf25Y39T04" rel="noopener" style="color:#237b94" target="_blank">1,600 tools</a>, which can include virtually any resource that can be used in an experiment. These can be software applications like search, but they could be robots, lab hardware, or human technicians. The authors hope to create a standard for all tools used in any experiment.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Behind the news:</span> SCP draws on earlier data management efforts for generalist AI agents and scientific inquiry. It extends MCP by enforcing tighter security, managing experiments, and providing specialized drivers for scientific tools. It also builds on earlier protocols for scientific research, including <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lSW3Dgjdk8FNWHjW1PxSLM7dg6LzW3sxSZv2mmGCgW7LbwFV7768qyW4XJ0jP3n6ZfdW8gn6mK3kKgdgW7fPr7P5-S6sjW1Q5js_4VJ0GrN3Jd6XZh8cKMW8RkRBx370_FDW2pN_K86vBKg3W6HJ6QH2tb_YsW3wVcG079gmvkW3XzkqB6ZjWGhW6GLH5T1kktN5W6kTXH76LFJnKW5fNpgC6C9LQJW1SzVDp6Y43fbW2r5c7c8Z7QGQW2lwXd44Sd4VbW8x49k48jHH_PW2-CKTC3bzwqlN3HYmD99HVM_W81YPbj4mpk6MW1JNxtV6_jBC-N2cnzT38hs99f1s6m-R04" rel="noopener" style="color:#237b94" target="_blank">A-Lab</a> (materials science), <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3lcW5q1Yj81vV_dCW6KZxR82V0jqrN3bhbd2GK2JsW2J77Yg8_7J65W8c0WTP6M05vKW4QF6zm6xxqZXW3V-n2y6BnLDhW6QWLJL2TKbQZW6GW8815tBZVMW4kBv8d21gvrqW1RlGC999knb7W4RdPyJ7tJH0BW7-SCb477-zvWW3ytWWY8XbkxTW1kHjD754FtBrW5VMdgQ9cwCdsW53R5RH2GlgvLVkDsRX5FNw8FW81fQZV5G2fy5W8rCs7m8fn6GWW2LT7FX4GMDHCW3pb9T97NPGY2W8lYD9x33t1J5W5DZgNM2FPccwW56S20W5MZ2hSW2r85sT6SLpD8W85-HXH4QSdCNW5jrDrV9cy6NGf6-t6NP04" rel="noopener" style="color:#237b94" target="_blank">OriGene</a> (biology), LLM-based approaches like <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lwW48nFzG92xhjdW2pnlpr8CsN2XW6dQgPl25_yWJVg30-b6VQRqhW1Xg17b1KzjRrTtmdV233_ZSW321pH17RHHx2N2rhhzmfVVW6W2qfk7d7149V_V9P8_p4BJbZbW4rj9ms8K2xFtVKyRJc7WTTC1W4TmSx5736gzxW887rmw2Xq9j0W8kjZcn8J4szXW3XWYhc76fdxzW1zKL353hMJ2BV8lxxR3WlnrLW64y7pX5d_RsNVK7DWy8vv4HdW5ZvrgW6T2wvtW5613Wc1fkfMnW6c34J16myc7-W8-HxYX5tbHVmf3L1-Vz04" rel="noopener" style="color:#237b94" target="_blank">Agent Laboratory</a>, and agents for specific tasks like <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p2VsQdl84n_5V1W12nSNt5RQ3NtW8M1Zxp18QryKW28ld_f10tJGrW1Jl4g18VynpfW809S7c3nt8WdW3Jzczs4qYKJrW1PJ2Hk5rBCwZW6CrJLB13Fy-dVWq8W28Vpx_jW5l9vQZ5FdG5DVpbNvk8CtvtgN2vLg-WmKzvbW8B4tyN1-7KCXVfqwbQ8csBVkW9gn9hc8rGX14W1XHTYQ2dG5H6N2LTqblF0HcXW24CT795CLvJxW5zyC5F6dHq87W4Jg3GP32bfv8W82l23p3Smh9tW1HHrHl1hSFQSW785k337FmP39W5Vmv2m3RyQ7qW1Jrtcd7TBfNNW8mpzb_5mpXQWW71JBtb4CrDWlW4nlDbD6TlZ-tW3c-31L4CFXHHW29qyC08l6gFSW1dpHM980qMTvf4s_Mcq04" rel="noopener" style="color:#237b94" target="_blank">Biomni</a> (biology hypotheses and analysis). SCP, however, aims to be more general than these field- or tool-specific resources, allowing researchers in a variety of scientific fields to standardize their methods and better foster multidisciplinary work.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">Why it matters:</span> Scientific research relies on both human and technology working in concert. SCP aims to standardize the connections between them. It can manage both simulated experiments that use only computing resources as well as physical ones that involve robots and other lab equipment. It also allows for better communication between institutions and disciplines by supporting dedicated servers on bigger networks. These distinctions (human/robot, digital/physical, disciplinary differences) are beginning to blur. SCP is a step toward that future.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">We’re thinking:</span> AI is poised to vastly accelerate scientific research. SCP offers a standardized way to connect specialized models, like <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3nHW7Z994S45dCSjW6VMBD15Yps97W5ZGq8_2KYWwVW5KYdRX7llCcqW31FF8w21VJStW6SNnGN3mpjNGW2P_vcV2yP-KzW1sTMc63WkPzjW54XxGL8h8xYnW5dJ4V13z23gWW8F2_t28Y40YPW5Xsfjr8FvJwZW2StFKj6gD24kW1zT17j1spJG2W5YxGdM2T_CpVW3m1LQN5SRL41W3QnhlD4jfX1PW4xgSwd99pNxLW5gH_1-2-CxP-W15QYqh1mt7GZN4YrDSKRQxbZW96kWbF5L1WttW8Xs7L71hgn6nW2ds8Cg1L1YSrW1bYmlt5H1wXvW7MXJsh625TQKW50MsT07shBYxW6qwyZV2KJcNyW3zzdX84PMxlBW5VKYfS4v9t2ZW6wz2qq3Q85pmW3YJyRF4N_5Pwf37V1vM04" rel="noopener" style="color:#237b94" target="_blank">AlphaFold</a>, with systems that automatically generate hypotheses, such as <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3l0W8gG20m8KmqwlW9gT8bk6ywHzkW65q5vv1XvQ8tW8yKcG07lRFNlW5RgSFp68xlVlW7NZDzh8cDVjlV_cB0z9g_hRnN7Ml713XKkv1W7pRHD37qSqLpW4ZZDR45DW7zQW7L5xHh14-QR9W4XkcyV57KyW2VM9TSw3Mc7DRW1SV-Gn8Ftx15VhtbMK4YRMwJW8WSSTt5Gv9HQW8Y5MZM564qBDW8l61n_3CScnyN1kTJHG_qSC8W1WbBw-6m-2CqVN9-V62jcTJTW41kCc3462CrmW7bMQJ_7JPzQbW93kmlb5lzwM0W87mhbg6r1qLVN9jT8TmHvc90W1pXHK079r-DlW76Xg0-7JjtJHW6SKLmp8pzNwLW80CLkk2Lm6CKW5ZZ9Cf75q-W9W7HNFZK5JzWZZW2lnKt73Dbxg4W9j3TWy281b8zf5HRk5g04" rel="noopener" style="color:#237b94" target="_blank">AI Co-scientist</a>, and robotic labs that test them, such as <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3plVzpJ1N5DfrxvW81pm1M17YNnWW3nYxqw4kV8-9W3trPFb5sn5CCN6LHddMrBRHwVXfnGq81GtVgW2X-bL340pmKHW5W72zp94dMpnW52bwYS89_4Q0VDyHPw1ml3qrW5XGBQZ40Sj0yVmgvnk6PbfJ2W8zN8nN8qvJbGW2vzg9R3bR_dmW4dxl0w1-_z3kW9brgnG8XMB8dW8GLjbD2R88yKW5h6ND74vRZP7W9c0Bsh5dvhywN2-FrNxn2nXzW6y_0Hg5P_1XZW99YVrG8WvD0GV2jkzK7RXvD-W1rsqnH3_3D0BW78xHTG7S-wZVW18_2-x6sXQQ-W8ZxhT82tc41TW3jPPcW96pfzRW3xC4Xg6M4hLXW5Sh8nw6Y0nzFW5qNm9-1TYc87W5yR5Zn1gD1rYW6-zgvh7GnX69W84HjTS4f_TH4f6Xtm9604" rel="noopener" style="color:#237b94" target="_blank">RoboChem</a>. This automated experimental workflow has the potential to advance scientific discovery at machine speed.</p> <p style="line-height:150%"> </p></div></div> <table role="presentation" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old13" style="color:inherit;font-size:inherit;line-height:inherit"><table role="none" border="0" cellpadding="0" cellspacing="0" style="vertical-align:top" width="100%"> <tbody><tr> <td align="center" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word"> <p style="font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto"></p> </td> </tr> </tbody></table></div></td></tr></tbody></table> <div id="m_-6958527957600471347hs_cos_wrapper_module_17241662744162" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px"> <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3pqW5scJjf7Jw5KJW1VnRMw10zC2WW8s9SYF15VmxJW21G3Kd8DrDmhW4kM9G_5Z-clDW7Wm7k08Ys9WNW173Ym_10b0J3VTNtnh8mkygMVQV9dQ652DvsW8_nNrB7_ZQ4FW6bJsWb1JDQ4jVZl1yB3mGK27W34H3Mt8PVL40W6TYwZY3xL1jYW96rYvp5RnQ6cN9gWSkdXKDw0W7nf2rr2C2tPNVhmpcT2SS1FGW6_q53h8NnnDMW5dQWL66-DGxfW2TgDZr8PpdDVW8_whrg8Kdcg2V_ZJ9s5zv73BW2vQn9S47sZCRW35-nr42F63HfW5wNhfh4lKcblW44Hq5R2f3dZfW5K_zxC5Z1pXRW6RkYjd2Ypm1SW82wF0q86WChFW2hRRVL8q60lnW5Zk7rK5h5GkcW3C3Q-K1Zl6-6W1w2sVW8XctJwW8cnmM82m6pWGW51zNNK296f3TVZS86G82w7zmW1q_vvr34CnFkVhvDBy1_SNQxW4Jzlwc9lccBSN4wrC2wpK84JW3K0B-j5m_ckPW4pGDDS4wjP-sW6tmH4c8DXjQJN78BQrr1f1CQVmPL7P15DxrTW2NWFCY5YL3rvW1K8cf5325nNjW22-BMy8qdPrKW8GbW6v2-RXCSN5Wvn0FjqfPhW4j_26-7SgNFPW200c5n8h0Qd7W3v2_JZ3vfY7SW6Yjfj659_rzKW8QPGrZ8hzdfJW2f-Dyd4jqT1xW8wwsB14vWBtKW7w3xdm1JwmLmW4Z8LH23rFhqDN4nxpBpXpzJCW449snq9342ZlW1Gtjw48nzVZ5W60mXVb5PB_lJW3LX2-38F8tPjVvSvQw6G6W5PW3BmXy0489yTvN1hBh61DNp6pV3S58r69g-1lW4h1Bgr8SpyN6f2gH59K04" style="color:#00a4bd" target="_blank"> <img alt="Scientists in lab coats work at computers analyzing molecular data in a modern laboratory with diverse team." src="https://info.deeplearning.ai/hs-fs/hubfs/image%20(51).png?width=1200&upscale=true&name=image%20(51).png" style="outline:none;text-decoration:none;border:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </a> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old24" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old24_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold">Learn More About AI with Data Points!</h1> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-family:Helvetica,Arial,sans-serif">AI is moving faster than ever. <em>Data Points</em> helps you make sense of it just as fast. <em>Data Points</em> arrives in your inbox twice a week with six brief news stories. This week, we covered Meta’s $2B acquisition of Manus to bring autonomous AI agents across its platforms, and Google DeepMind’s partnership with Boston Dynamics to deploy Gemini-powered robots in real-world industrial settings. <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3ljVhhS-L5yD9kBW2_G34Y46gCJdW4HtZFw7Q12hQW6gQnJY4BKd47W2pMgm_8YClbcW7MYYCF2YgnGzW3qtdY58wjZwdW933bw49hBT71W1T4ZXB1xVpFwW8Zq7my1cVg6RW50FZRK2JdyBpN8BFjb22r9lXW3DTnjb1YgVTYW4N034F22m6hxW5j-X-V4MhNybVpW58w97T0nCW6bNXJ544b181W4qQHtk7gGfK3W3KgFqm3nwzBvN865vysq95D8W5gDVZy1Dd0BPW5mh-tm1V0SXKW2Wb-PB8S1C0TW3gr84C6PrgwXW42lfh07rMN7kW8m6l9b8yVy7QW6YYfBl7FH6vvW2pn0b11hMVf2W4dwxmj49D5_KW1DJsds3RvbCQW4DDjqS8nBkmCW8S_zKw1Yk4LSW2MlMJX1Wl12KW6933tP8CVryvW48zvMv2ZzR3wW3vLMNB5PNDvhW57s6hb2Y-d49W5K089b49j6NfN9dDGthF6lwjN4CNMXK2lqwbW23lt166dnC92VqqbXS3GjBBCW4vBShf6n-D3dW4X_j-B1HJJGmMGCpC9ZPr4gN8nSslPp337bW81sBjl8FZ-ZzW8R3V4P2K1lVPTMCTv8mF3yMW3G4cqK4sb7KvW4WKhJV53NskwVP4dXr4w3nH2W8F85BD2872SgW88nzBk2b51t1W3jtzNj9bYGPJW7VzT784RtYVZW7Cdl4v9fJtCjMdpjf6FKmJLW5L_qsZ74r2tlN6yTmgCs4NTyW5l0jq44SN6sZW44-9BN2mX5qDW8ZWf991VFDCTW8FXLPw49m91XW42wzlV4Gs_YYN1Ys5QkK3nBNW7ZPnhX7rnfdfW8dwlDR5GQYjXW8lCs0v1_qLS_W9hxh1M1B8mTsf26kcrF04" rel="noopener" style="color:#237b94" target="_blank">Subscribe today</a>!</span></p> <p style="line-height:150%"> </p></div></div> <table role="presentation" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px"><div id="m_-6958527957600471347hs_cos_wrapper_module_17370394853571" style="color:inherit;font-size:inherit;line-height:inherit"><table role="none" border="0" cellpadding="0" cellspacing="0" style="vertical-align:top" width="100%"> <tbody><tr> <td align="center" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word"> <p style="font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto"></p> </td> </tr> </tbody></table></div></td></tr></tbody></table> <div id="m_-6958527957600471347hs_cos_wrapper_module_17370394878062" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px"> <img alt="Graph with 10 colored lines shows topic ranks monthly, based on a Microsoft study of Copilot usage." src="https://info.deeplearning.ai/hs-fs/hubfs/COPILOTUSAGE.png?width=1200&upscale=true&name=COPILOTUSAGE.png" style="outline:none;text-decoration:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old21" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old21_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold">Copilot’s Users Change Hour to Hour</h1> <p style="line-height:150%"> </p> <p style="line-height:150%">What do users want from AI? The answer depends on when and how they use it, a new study shows.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">What’s new:</span> A Microsoft <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3mbW4dLdB5378lJ_W6BM3tg55SDZdVbHsPb10BCX3W3fYZHv4Xw4pdW6fRGCy8slM_lW5r8Bg91kF-2hW1y36L18wZyG3W6pJG3Y5XVlXjW4vsZHs4b2stbN7rbthn4nwCrW1XM0cX4YG2T4W6Rk-kq75_CdzW7cqVZx2NTTvCW6HPgdf3mF_rxV9TVtd2lyZL4W8YKF9h3hHpxYW20Wsg81yknNpW1dYQ4y6g8q_0W2rkL__2qrBLdN774pLxVh5YhW179RlC1lyT1CW4S9pmy5fW3mbW8cTBV46r75ckW38NDtf12RYsPW4qJ0kx4BBjCdW34CkJ11z7CK-W47S6kl93WJ8sW4Fw4964FmQfvf4svdCC04" rel="noopener" style="color:#237b94" target="_blank">study</a> reveals that people used Copilot differently late at night on their phones than during the workday on their laptops. Conversations that focused on productivity and career were more likely during the day and on desktop devices, and health, gaming, and philosophical questions dominated non-work conversations. As 2025 went on, more users asked the AI agent for personal advice.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">How it works:</span> Researchers <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3kSW3_s33w6nj_V5W1NsHyQ16gdWYW7FsRTD7d0WydW41QJ8J4fRfk9N3ptYklrjwq3N7MRQD3PWPy2W3JgDy-53HBKQVgd_z97LDn6PW4DXzYT1678thW5qWlPB1HH1Y-W3SRzGS1HjyqxW6z-Lw480bPc0W2F-yv98JmTHsW5kJBSg2WqnSNT8v7175Rdf-W28ngBM3zbtNsW63MJKw8fvpQbW8dd7xm2Lh7bhW8hhhqK1DPtNlW4zlr2-8lTDHFVV-cVW1LkLKdW1gQhPL5wLzPCVpcjYG1pwDhYN2tVwmsZQ57pW4_z5h26QP7_gW4TgYVy6gml4gVgK4hl6fqQfzVYxSvP47T8d5W4P5ZRN2B-s9VW4kgqff6nkTWVf6GCBcF04" rel="noopener" style="color:#237b94" target="_blank">analyzed</a> anonymized summaries of 37.5 million Copilot conversations between January and September 2025 to study how customers used the system, making this the largest study of its kind to date. The authors conclude that AI has become more socially integrated, as users employ it in aspects of their lives beyond work.</p> <ul style="line-height:150%"> <li aria-level="1">The authors examined a random sample of Copilot conversations by paid and unpaid users, excluding commercial, enterprise, and education accounts. Each conversation included timestamps and device type. The authors used AI tools to summarize roughly 144,000 conversations daily. They built classifiers to assign each summary a topic (like “technology”) and intent (like “seeking advice”), identifying about 300 topic-intent pairs.</li> <li aria-level="1">The study ranks the frequency of topics and intents by time of year, time of day, and device type. The top 5 topics in order were (i) technology, (ii) work and career, (iii) health and fitness, (iv) language learning and translation, and (v) society, culture, and history. The top intents were (i) searching, (ii) seeking advice, (iii) creating, (iv) learning, and (v) technical support.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Analysis:</span> Topics and intents differed depending on device used, time of day, and time of year.</p> <ul style="line-height:150%"> <li aria-level="1">Users were much more likely to discuss health and fitness on mobile devices than desktops. Seeking advice about personal matters spiked near Valentine’s Day. Philosophical questions became more common later at night, while entertainment-related conversations plummeted during the workday.</li> <li aria-level="1">As the year progressed, topics and intents became less focused on work and technology and drifted towards social and personal matters. This shift suggested that the user base became both larger and less technical and/or users began using AI for both personal and professional matters.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Behind the news:</span> Microsoft’s report follows similar studies by some of its AI rivals. </p> <ul style="line-height:150%"> <li aria-level="1">In September 2025, OpenAI and Harvard released a <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lRW37hpZn1mJq4NN3YRn_qB0gp0W3K-g-h8lhHdCW1GpTqN7FyWflW15dtnt4r-qWQW8-rrXJ8XM6-GW5gkrNP3G2dX1V2y34w54Wwy8W60QD2_9b5kP5W4Mt5W21N4q9tW7Km6kG2pjpZPW9f0-GR6_C3RLVYd6LC7LGhRvW7YmfCp1_xcTSW4JNMr33bsLpjW7mdglL569p2xMssSLKZ_Wc7W65ljLW3Wrl1hW84XCly7hSxHZW87CF0z4lwQxMW6v2Nw-7xR94tW6DGHVQ2jn857W7FPV_q76RMzsW99mxs_4fGXMvW8cg52j8d1hDpW5MvrJn5RcXcGW83glV05l-5mRW4r2KTj1xCWnNW49Bzmy7XqtMWW2Y3qF47bRPXDW6nHx592MpnKpW73y5NL1LcMz7W49HGfZ28RJFtW61Cpqp5C0Rjqf1CKkLY04" rel="noopener" style="color:#237b94" target="_blank">study of ChatGPT use</a> from 2022 to 2025. It showed that 30 percent of uses were related to work while 70 percent were related to non-work activities. In addition, the gender gap among users among users shrank steadily during that period.</li> <li aria-level="1">In January 2025, <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3n2W3FZwH72gQLPmMDFTCMKTk4JW8Sp0Dm8m1h3NW8RH7W_3V1lL5W4lHG7X5MX36_W23hmkW5t954HN1DfYFhV0TkBVh4gP329XMV9W82F4JW3c8qCmW1wfyqy4z-kVtW44zxzn8cl1ZhW4by7Yh8_szdzW4hP3M07VxSjHW4nQczv18c48BW1ss5yQ59yckrW760lpp6WW-5lW2lhgnj292JLdW6zZ1yZ7Bhg3WMJRhrdCk4LdW7dzMVK1Q-mkKW32N-Mb3SXM5fW43YmQ560pSdtW8x1HFj8L6k4xW2dSKHF4pBNSDMGr_3q7z1ZyW5GPGcj98lTFNW1vDbcF9gjmwdW735N-b2VC6QtW4hk2sF2XNm9SVT9Z235BP9zCW3RPKBt1xs5zsW4DhHvy5JMT3Ff4FzF6404" rel="noopener" style="color:#237b94" target="_blank">Anthropic’s study of Claude</a> showed that the model’s user base focused on work, especially software development and text communications. A small but growing number of users engaged in games like Dungeons & Dragons and sexual roleplay (despite prohibition of that use by Claude’s terms of service).</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Why it matters:</span> The authors argue that the AI community may need to rethink chatbot design altogether. If users treat chatbots differently on mobile and desktop devices, AI builders would do well to design their systems to suit the devices that will deliver them. Application design is one way to accomplish this, but system prompts may be another. Desktop chatbots and agents can respond with more information-dense answers, guiding users to execute tasks, while mobile agents can offer shorter, more empathetic responses.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">We’re thinking:</span> Studies of chatbot usage conducted by different companies show different results. Perhaps each company’s users treat AI differently, so the results of any given study may not apply generally. That said, the Microsoft study suggests that the device used and the time when it’s used can have a big impact on what users want — important considerations for designing any application.</p> <p style="line-height:150%"> </p></div></div> <table role="presentation" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old22" style="color:inherit;font-size:inherit;line-height:inherit"><table role="none" border="0" cellpadding="0" cellspacing="0" style="vertical-align:top" width="100%"> <tbody><tr> <td align="center" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word"> <p style="font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto"></p> </td> </tr> </tbody></table></div></td></tr></tbody></table> <div id="m_-6958527957600471347hs_cos_wrapper_module_17241662955853" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px"> <img alt="Diagrams comparing LongCoT and Delethink environments show reasoning processes and context management." src="https://info.deeplearning.ai/hs-fs/hubfs/MARKOVIAN.png?width=1200&upscale=true&name=MARKOVIAN.png" style="outline:none;text-decoration:none;max-width:100%;font-size:16px" width="600" align="middle" class="m_-6958527957600471347stretch-on-mobile"> </td> </tr> </tbody> </table></div> <div id="m_-6958527957600471347hs_cos_wrapper_module_17370395015403" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_module_17370395015403_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold">More Affordable Reasoning</h1> <p style="line-height:150%"> </p> <p style="line-height:150%">One way to improve a reasoning model’s performance is to let it produce a longer chain of thought. However, attending to ever-longer contexts can become expensive, and making that attention more efficient requires changes to a model’s architecture. Researchers proposed a way to limit the cost of processing long chains of thought with just a bit of training.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">What’s new:</span> <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nKW8SXyhJ7DQcqSW8nHYDW27ZnkMW7-qqxL1DZ4LXW99kBsN2HH6cXN3hSL0PzN9_GW88SD2t62Xn-NW7jC2pZ5GwK4CW5NRxGv8YSxJpW58GnTd4-W-PbW4VG01n68vhq3W3c4Hvz3trwNJN5cdb0q5KsgxW6q-32N7fK-0lV4JBTF5qrtGGW9jgDm36dLyBMW91y9R58pXr96W69J70p7dz_LvVbq9_b5gbqjmW12FCCy3YdKTPMWv2gLQgnBKW2Qvm1-1VWPlsW8731bg6JnDL0N3Ry2sX9pj82W2l-hmn1M1NwDf8Qx6z404" rel="noopener" style="color:#237b94" target="_blank">Delethink</a> is a reinforcement learning (RL) method that trains large language models to periodically truncate reasoning tokens to a fixed maximum number. The authors include Milad Aghajohari, Kamran Chitsaz, Amirhossein Kazemnejad, and colleagues at Mila, Microsoft, McGill University, ServiceNow Research, Polytechnique Montréal, and Université de Montréal.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">Key insight:</span> Reasoning tokens typically accumulate within a large language model’s context window, where they consume quadratically more computation as the contents of the window expand. One way to counter this effect is to train the model to reason within a maximum context window size. In effect, as a model is reasoning, it can learn to replace its chain of thought periodically with its latest “thoughts” and then continue.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">How it works:</span> The authors fine-tuned R1-Distill 1.5B, a large language model, on math problems in the <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHq25nR3bW69t95C6lZ3mTW80Q7xZ4QXs__N1lxhWpcckqyW8wVFFy6hzJcVW7lSDK17bJLCDW4lhZLT5Sd68CW9h4NJR60THQ-W5k4NRL8Q42SYW80l_MJ80xtntW6wHP2T7R1_QvM3MrcZJQBj4W7sKMf89hjZJZN7mGhVXYlC5JW75Xhct8CB4X6W48SvnV7pRGdlW3LDlbC5mZnl1W1lM38N4LctGlW4Kvmxd7_MBlFW2D9djG45RMj6V3PgnB349h89VDJ0N86CN-X-W91qhB675mcPDW3Xv7FR3zltJCN6mtqCgyZ0QYW1h0rpD6KVCZYW2lzJK27Fp93jW3rtB8P6p3RQbVT5mqm8LrDz8W4-kgsq6p-6fTW3MytCj3R3WLpW38hspr2T_sqRVXXhZ_3fxpH4W1s9Xp_36n6YCW9681k46Q2Cn-W804H7f3-t3N3W4ld37c4JpVXDW6dp7BY8lg1Z5f3ztR6-04" rel="noopener" style="color:#237b94" target="_blank">DeepScaleR dataset</a>. They used a modified version of the reinforcement learning algorithm GRPO that trained the model to reason in 4,000-token chunks:</p> <ul style="line-height:150%"> <li aria-level="1">Given a math problem, the model generated a chain of thought until it had either finished or filled the model’s context window with 8,000 tokens.</li> <li aria-level="1">If it didn’t finish its chain of thought, the authors replaced the context with the original query plus the last 4,000 tokens. Then the model continued to generate its chain of thought until it had either finished or the context window once again held 8,000 tokens.</li> <li aria-level="1">They repeated this process until the model had either finished its chain of thought or produced 24,000 reasoning tokens.</li> <li aria-level="1">Then the model attempted to solve the problem, receiving a reward for a correct solution.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Results:</span> The authors compared their R1-Distill 1.5B models to the same model after fine-tuning on the same 24,000-token reasoning budget via using GRPO. They tested the models on reasoning budgets of 24,000, 96,000, and 128,000 tokens.</p> <ul style="line-height:150%"> <li aria-level="1">With a budget of 24,000 tokens, their model matched or surpassed the baseline on all 3 math benchmarks tested. For example, on AIME 2025, Delethink (31 percent accuracy) outperformed the baseline (29 percent accuracy).</li> <li aria-level="1">Their model’s performance continued to improve as the authors increased the reasoning budget, while the baseline achieved much smaller gains. For instance, with a budget of 128,000 tokens, their model achieved 35 percent accuracy, while the baseline achieved 30 percent accuracy.</li> <li aria-level="1">The authors estimated that training their model with a 96,000-token reasoning budget would cost 7 H100-months, while the baseline would require 27 H100-months.</li> </ul> <p style="line-height:150%"><span style="font-weight:bold">Why it matters:</span> This work eases the quadratic compute barrier that can make extremely long reasoning computationally infeasible. While other methods, like linear attention, achieve the same result by changing the attention mechanism, Delethink restructures the reasoning process to limit processing regardless of a model’s attention mechanism. It opens a path to reason efficiently over longer contexts without requiring new model architectures.</p> <p style="line-height:150%"> </p> <p style="line-height:150%"><span style="font-weight:bold">We’re thinking:</span> As the authors mention, most LLMs are pretrained using relatively short contexts. For example, Llama 3 models started pretraining with examples of 8,000 tokens. This may have made them good at processing inputs around 8,000 tokens long. That is to say, Delethink’s performance may have been helped by the fact that LLMs tend to be pretrained on short-context tasks.</p> <p style="line-height:150%"> </p></div></div> <table role="presentation" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old25" style="color:inherit;font-size:inherit;line-height:inherit"><table role="none" border="0" cellpadding="0" cellspacing="0" style="vertical-align:top" width="100%"> <tbody><tr> <td align="center" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word"> <p style="font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto"></p> </td> </tr> </tbody></table></div></td></tr></tbody></table> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old29" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old29_" style="color:inherit;font-size:inherit;line-height:inherit"><h1 style="margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px">Work With Andrew Ng</h1> <p style="line-height:125%"> </p> <p style="line-height:150%">Join the teams that are bringing AI to the world! Check out job openings at <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3p5N65swpFxLdcBW8s0KY56mqGLPW54Gqb655n5yLW2VlKw93KfzxyVsN0Xb6JKmjBW58N5lm172PBQW1pyv7s2nLPmJW425jTD7Kg0ltW6L4phx2-p_-fW4f56hZ8KXv_QW92xSwf4_ytncW4592Ww3PryZ1W8573kw11T8gtW11NgFR1dbB96W3YbCdl4xQrpNN81SyzgkdP_TW3X5ZmK4ZtpbnW8SFYXH5z_h9FW56-zy35zYthhW7Z8hQ14HdSpGW86N-fM4Fr3J2W6sVzYm75ZmQDW1yYb9B41ckycVXn2j82_vpHcf6T8cZl04" rel="noopener" style="color:#237b94" target="_blank">DeepLearning.AI</a>, <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lbW49jm1g3GPY-6N6cnvR7rjkr-W8VgP9N5JN8qsW6yShD98ZcZqKVrKLqv3dPMCpW718j8y5JdRKMW2tsL9K4nVRPnW686Cl16p2Bc4V9V3rl94vpRvVm48ql1qZpmxW68CxMZ5CfhzDV5KS6l8P2GyXVWcLT93KYJZcW4cs8CQ4dWkV_W8-rzXF4LCJjWV7Ltx08PgvF6W3ZKBYn7LjkykW4w4bvB4b9dgpW5q3QZW987twRF903KzsnMM2W7r5RZ12C1vQWW6bZJhw9htKGMW8lFfhm2DbYx_W7-NxNX7GdCfMf3Z3kkT04" rel="noopener" style="color:#237b94" target="_blank">AI Fund</a>, and <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nTW1lpkY37j3bCYW33RpBy4V29pwW98gl3t4BfKvGW4zYqFT5Qjxl6N1RXgGSVm1WrW7xP5T35MzQ_LW8rWqPX5ZW-dKW4DVSpQ1S4YbXW392rhC2-v-X3W6Fs7xg7RtQVdW9llkVF7-jWgmW2Py9Qd71P794W77Js7c6wGQy9W1mDLHH4MFp59W7FbMRR989XsQW97sLMj6YmLRbW3GXV_R3XnjFyW2qLwG01d0fwXVbzPfm2hWP9hW2Y8Htq6wMdHkW8PqLSG2S1-NjW8G4rfc5dtp1SW8jKk_Z44jcmMW6WBJ4K1Z0S1_dZvf6804" rel="noopener" style="color:#237b94" target="_blank">Landing AI</a>.</p> <p style="line-height:150%"> </p></div></div> <table role="presentation" cellpadding="0" cellspacing="0" width="100%"><tbody><tr><td style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old30" style="color:inherit;font-size:inherit;line-height:inherit"><table role="none" border="0" cellpadding="0" cellspacing="0" style="vertical-align:top" width="100%"> <tbody><tr> <td align="center" style="font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word"> <p style="font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto"></p> </td> </tr> </tbody></table></div></td></tr></tbody></table> <div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old31" style="color:inherit;font-size:inherit;line-height:inherit"><div id="m_-6958527957600471347hs_cos_wrapper_hs_email_body_old31_" style="color:inherit;font-size:inherit;line-height:inherit"><p style="line-height:125%;font-size:14px;text-align:center" align="center">Subscribe and view previous issues <a href="https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3prW7Qb8g-2wx0YNVy8LrQ4CN6dbW3MHBx77698fhW8QshbW4N0vzqN4Prl3f180wZW1MF8S644trRSW1SwZL88hTcWfW2t4t953T985BW7vH8P03bfd35W86y-3F1K69nqW8Tg_6s4rlkY0W5_7hHx7fv5-RW1rtZ032W9C_xW3-NFd_79zfRNW1WgY_z5R8lQzW7YzSr-2J5t3TW59mPLM72YyqMW2XhG3v2PBjKcN2Bj7GGwPWv6MYndxzMw1qGW1N2Vmb1Qr12jW51_Yxt7LG-W1W1tjyPX7mh_KhW7LJPvS60H8jDf1gjZhP04" rel="noopener" style="color:#237b94" target="_blank">here</a>.</p> <p style="line-height:125%;font-size:14px;text-align:center" align="center"> </p> <p style="line-height:125%;font-size:14px;text-align:center" align="center">Thoughts, suggestions, feedback? Please send to <a href="mailto:thebatch@deeplearning.ai?subject=RE:%20Thoughts,%20Suggestions,%20feedback" rel="noopener" style="color:#237b94" target="_blank">thebatch@deeplearning.ai</a>. Avoid our newsletter ending up in your spam folder by adding our email address to your contacts list.</p> <p style="line-height:125%;font-size:14px;text-align:center" align="center"> </p></div></div> <div id="m_-6958527957600471347hs_cos_wrapper_footer" style="color:inherit;font-size:inherit;line-height:inherit"><table role="presentation" width="100%" cellpadding="0" cellspacing="0" style="font-family:Arial,sans-serif;font-size:12px;line-height:135%;color:#23496d;margin-bottom:0;padding:0"> <tbody> <tr> <td align="center" valign="top" style="font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;text-align:center;margin-bottom:0;line-height:135%;padding:10px 20px"> <p style="font-family:Helvetica,Arial,sans-serif;font-size:14px;font-weight:normal;text-decoration:none;font-style:normal;color:#3b3b3b"> DeepLearning.AI, 400 Castro St., Suite 600, Mountain View, CA 94041, United States </p> <p> <a href="https://info.deeplearning.ai/hs/preferences-center/en/direct?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159" style="font-family:Helvetica,Arial,sans-serif;font-size:10px;color:#3b3b3b;font-weight:normal;text-decoration:underline;font-style:normal" target="_blank">Unsubscribe</a> <a href="https://info.deeplearning.ai/hs/preferences-center/en/page?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159" style="font-family:Helvetica,Arial,sans-serif;font-size:10px;color:#3b3b3b;font-weight:normal;text-decoration:underline;font-style:normal" target="_blank">Manage preferences</a> </p> </td> </tr> </tbody> </table></div> </div> </div> </div></div> </td> </tr> </tbody></table> </div> <img src="https://info.deeplearning.ai/e3t/Cto/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsW6wcH9r3Cy50v122" alt="" width="1" height="1" border="0" style="display:none!important;min-height:1px!important;width:1px!important;border-width:0!important;margin-top:0!important;margin-bottom:0!important;margin-right:0!important;margin-left:0!important;padding-top:0!important;padding-bottom:0!important;padding-right:0!important;padding-left:0!important"></div></div></div><div><br clear="all"></div><div><br></div><span class="gmail_signature_prefix">-- </span><br><div dir="ltr" class="gmail_signature" data-smartmail="gmail_signature"><div dir="ltr"><div><div dir="ltr"><div>Markus Edgar Hormeß<br>Founding Partner<br></div><div><br></div><div>WorkPlayExperience GbR<br>Oedenberger Weg 14<br>D90571 Schwaig b. Nürnberg<br>Germany<br><br>+49 151 24101032<br><a href="mailto:markus@workplayexperience.com" target="_blank">markus@workplayexperience.com</a><br><a href="http://www.workplayexperience.com" target="_blank">www.workplayexperience.com</a></div></div></div></div></div></div>Full payload
{ "event_type": "email_inbound", "timestamp": "2026-01-09T12:34:17.425361+00:00", "swg_uid": "04-a4020826-479c-4970-942d-a46bc59c645a", "from_": { "email": "markus@workplayexperience.com", "name": "Markus Edgar Hormess" }, "to": [ { "email": "receipts@inbox.workplayexperience.com", "name": "" } ], "cc": [], "text": "---------- Forwarded message ---------\nFrom: The Batch @ DeepLearning.AI <thebatch@deeplearning.ai>\nDate: Fri, 9 Jan 2026 at 12:03\nSubject: LLMs Go To Confession, Automated Scientific Research, What Copilot\nUsers Want, Reasoning For Less\nTo: <markus@workplayexperience.com>\n\n\nWe just launched a course that shows people who have never coded before, in\nless than 30 minutes, how to describe an idea for an app and build it using\nAI.\nView in browser\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx5nR3bW95jVnq6lZ3m9W4FHPhr8rhpmXN4bY_GGr2cxDW3JmQfL6MvqJ1W8N-Slr32QYRCW5tw1Fs8S11NRW6YwScS2ls0H0W8VmJJ92V6zZLN5kxtb_WssNwW22YZlN1vpFWgN2RYcnvfd9ScW1f6jc-5Cqm33W46NtVN59ZS0XW81-zG45dNvyDN7rmc1QGRg92W2nD_Bh47DC3yVxtVWs5JRV-JW965Pvs7Sp7WwW7lhHNF21ns0ZW4h2PSP94cCFWW3rNvJK8z0kf7VXQ95L55R3pgW3DC3tB5wR5mgW8GcYp18sMf0yW7PFNHD93ndzYW21KLx76Ft_txN77pF4pjPwtsN8k8yk1XtVc1W4rR-dX2XRN3nW76gTJ98sCM1CW6Hm1r-2nMFVhW1lxxdk8Zgts3W7D3ZFv9m0tFyW6Zpw0k7pKDsbW34ZSVn2nwNPFM3791TH8ZJXW8gTzff43q23ZW3dvKsz5TNcFlW5NYthl7qPq4pN3J0q-CrPjP1W44rHQ854vR2PW3-n0FD1Xcb6_V41BBx7Wz9L7W7mq1GH6Vz6NYW954P6513wyynW8nBMsK3pgtvMW3Bg8pQ6vsdgnf8B1frb04>\n[image: The Batch top banner - January 9, 2026]\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lvN4B5SvcbldLVW2KNgKQ4ZjL91W8HSh4212Cxb2W7yZy8K6JxxJYW3QJ4dw1-YPkxW6BPHMY2DgQ7PW4FPL221rbj_BW47BBdg6CqfcsW3yLgwB7r_K1HW6NXWtz2Dh6k5W99fcQp5YB-RcVnL-bJ3QcYWJW6FZTMB7NBhjhW1qhCfF6D6dy1V1nWkW5RJNtPW3jqW3G3Q8L-WVdq4217X2dljW3KnMvV98BR69M_KgHYgls0MW8yJKmj6_4yHDN3tJ2kt4q_9dW2GkMLj53rrz4W73zM0W4yNFKgW67NFZZ85lNBvf59sbLT04>\nSubscribe\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3mDW128n0H27S-49VngKyZ59WlPCW8rvm_06Js47MN26yq4ltTSgnW22Syvb4-FzZCW4yHs6L6ZZp4YMb4Wv4whLQsN20TjVNWnTNnW2pJBlK4NB0vKW84jDhw1jgxRXW4Wmr875hRCVzW2RdM7086dXRjW53GdyZ297BNDW7n1SZp60p-mJW4m_89W79tr8PW2vRF0G2wGTR1W4H8jL56R9vnGW5zBc6s8X6cZlW8rLTm07q3RDlW29SNyl6Hq3n5W83KlcS65zxrRW5kcrXd2rctX4N4SHqZGctbw6W7brsFB96DmjMf7hN02R04>\n Submit a tip <thebatch@deeplearning.ai?subject=RE%3A%20Tips%20and%20News>\n\n\n\n\n\nDear friends,\n\n\n\nWe just launched a course that shows people who have never coded before, in\nless than 30 minutes, how to describe an idea for an app and build it using\nAI. It is now time for everyone — marketers, product professionals,\noperations specialists, analysts, students — to build software applications\nwith AI!\n\n\nI’ve often spoken about why everyone should learn to code\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3n-W1mny-F56FdD0W1hcc0c7mP1dYW2TSZhj83lZK_W4p2y2J7hN1bPW3cShVp3pZ2rvN2K178TBd__BN6bKqT--9GL4W66HXht7VTMndW5kFyTL18ZSY7W6dNSqv4zhnxKW8mp-vV2bw2LvW5BHnk-3M_7GgW6KycxS677nwzW3_bRw-5Gpj5PN8wGfbFz_yFNW30ZZpw26vq1qV-xStt31rggtW2rxTHZ5TFrXGW60mvHB8k2ZDSW7HyXNM1WY5V0W4rKwhs8cmZF9W2s0pHq7ckmF0V7RW4c8hlN9DW6CYjqJ7-d-45W4SMhcK6dN3dcW4-RZ2C82dmv1W1WDmS12H8zM_W4xF-1Q43SRRVf8vfV3P04>.\nI’m seeing a rapidly growing productivity gap between people who know how\nto code and those who don’t. For many job roles I hire for, I now require\nat least basic coding knowledge. Many times, after I speak with a\nnon-technical audience about the importance of building software using AI,\npeople ask me how to get started. In the past, I didn’t have a great\nanswer. That motivated the DeepLearning.AI team to create “Build with Andrew\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3ndW1NFm7J4cS7HKW5-lT9S5G9KWQW3Kl8q36Lnm9qW8xBqQ91d9ltcW3c_ztd5Nj6fbW8_-wBW5lqVKxVJjGWn8gw6bjW50b5FY2SXBYrW6Rv3hn1hVDXbW9148Jv385H8XW1bDpHK8Dt0cJW2TGJF295ZhLmW5brfWH5JFWmMW100VHl1nR409W7njgC79jmb9GN2rkT_hQ9Xx0W3wZ87N3KgSP5W95mH222mxX_CW1PxCg475GFVQW2fqFFM236wR8W2Vm0M_7kzgyYW4R7Jtd6hFYdJW4gmbsW22Z7q-W4jf5r12-j1-SW7x1Spn3ht80GN3G989xnnzZ6f4-PBsH04>.”\nIt’s the best way for someone who wants to try vibe coding to get started!\n\n\n\nThis course requires no prior knowledge of AI or coding. And it’s\nvendor-agnostic. Specifically, learners can use these techniques with\nwhatever tool they’re most comfortable with (like ChatGPT, Gemini, Claude,\nor the chatbot built into the DeepLearning.AI platform).\n[image: A birthday card generator form shows fields filled with humorous\ndata and a chat bubble indicating help needed.]\n\nIf you take this course, you will build a working web application: a funny\ninteractive birthday message generator that runs in your browser and can be\nshared with friends. You’ll customize it by telling AI how you want it\nchanged, and tweak it until it works the way you want. By the end, you’ll\nhave a repeatable process you can apply to build a wide variety of\napplications.\n\n\nDeepLearning.AI’s mission is to empower everyone to build with AI. This\ncourse is just one of many steps in service of this mission.\nIf you are already a developer, please encourage your non-developer friends\nto try their hand at getting AI to code for them. Not only will this help\ntheir productivity, they will find it really fun as well. Please invite\nyour friends to come build with me\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lxW6Z-p6p842JrLW7yg_Fx7y0RthW4tZ2YR8bkSbFW2VWgNL3ppRTxW3pyCzV1rN1mJW3pDnwr4L9PHYW8TkLch3GsmnWW2Sv6rs87XlvjW5YMPmc5RG9KvW6LKz0C5t_D88W3CP8xy7d8LC1W7YsdXZ1cSFnnW44yGMF8g0xYBW9kqBQK4P9MjcW6G3sdy7KLffjW8VcD9_2lyJXHW5y_H-b6Y8N0fW6jWkhj1RykwWW8nt0-w6rqMNTW3944c41G6PG6W68_j_y3yL8bWW41llRh51L8lsW7ctSv-26xvF6VhRWQl3HPRLcVr9jjb3yjl4_W7crrh18zqYQSf6HXPMj04>\n!\n\n\n\nKeep building,\n\nAndrew\n\n\n\n\nA MESSAGE FROM DEEPLEARNING.AI\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql3qgz0W6N1vHY6lZ3pNN432hYGKZ9k9W65ntXq84NPj7W6QdjW-6jjJKTVSfVsG5bKFyLW5k_4LL1C0qc3W3jG5w38645YMW3Vwvdz5CMtqvW4wxhq82mXJtvMd26xqzx1jhW6C2qkT5CQH1YW7Kl5Zq8kkB3CW9b79PR7WLD_nW3_Z_Tv6-nJY9W6Ybjsj6qqdC5W7Tp1XG1WF_HFW2071jr6JqsxdW7JW4hH1HYc5mW5RZh736dJGS7W17sHWC6zkhK5N8klj4jtTH5PW5YJ0Yw4MQL-XN4FZn5dSyYQzf4XFBW404>\n[image: Promo banner for: \"Build with Andrew\"]\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3n_W2n5v537npkmlW34qHlf41HL6SW77FH7H6W9QttW8f8JCk6txg6HW5lww2h30YTYYMKnHkKTVbGjW3YFdWh1NFVc8W4sm7Mh9d4P3GW8cstBj6rGbTYN2gcvBhKGScpW7T8Vck8zsQqYN15mkhN19Y-_W1zbxz27-zjJtW7cvD9G57r7K7W5K_7pg5pt7XxW6LPKx_64ngJxW7yd5Fg1y1q5kW62qlvj5s1q0-W6lxjVL5RRS8BW7v6mX16g2Bh6W49w4tk2y1d-wVD8qp77KxYtcW7ChZSC8vk9FnW8ySxlv2Hm2bTVV4FT669k2xqW9jVRbS3XSdvTf4j1sxC04>\n\nYou don’t need to learn how to code to build an app. In “Build with Andrew,”\nAndrew Ng shows how to turn ideas you describe in natural language into\nworking web apps. Perfect for beginners, and easy to share with someone who\nhas been waiting to start. Explore the course now!\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3pmW8QHvqM4KhKGnW8T2m_K5z0fxnW6TsRVR2sJhynVX1TBS3RXnDhN4pD_vWBZ0ZtN1Zw0FJb7pQ-W6h2Vcc5mZp4XW550B0L8vlgd4V7F9G-2MzZ4fW8v-mZL7ZK5RLM83LrPP9wfkW6r6LWP1CBx9bW7n2Csx8FDDllW6chw1v64kfT5W3RPsXW8D8ytkN2wgWCy4kFpBW2dXNXg8B3m8fW9lFd_x2fmRG9W8fKcPb8j5d6JW48hDw56dxDGhW7FcD4Y5-MvCRW39lSTL66fcVNW9dkc8N329Yb5W2fXT0P5x06QhN3gyCFHPnlV1W1SFhf72rdZvff4r4PWM04>\n\n\nNews\n[image: Dialogue displays a model revealing it answered incorrectly and\nwrote code against instructions.]\nTeaching Models to Tell the Truth\n\n\n\nLarge language models occasionally conceal their failures to comply with\nconstraints they’ve been trained or prompted to observe. Researchers\ntrained an LLM to admit when it disobeyed.\n\n\n\nWhat’s new: Manas Joglekar, Jeremy Chen, Gabriel Wu and colleagues at\nOpenAI fine-tuned GPT-5 Thinking to confess\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3pZW7gKR221JKmCxVZ_gw13WSktRW2CwXWr8_dhr1W6yB4qg8VRbS5W2Ljd8m5RN0LkVJy-9y2NxCVTN539nfWKBvHCVwkKqN8636n7W4jThy671t1BNW3RPK5s3z9WbkW5n2JPd7sNF_cW5XcmT03ZMzYqW7qRklg37s8tWW8_-pz58N3mqgW5Ww9tg42dG5WW8fF8xm2mp2ZZW4LQC1Y3VTnX7W65Z3ZC4NqL_XM3NxWdVDkW3W8X8WmJ6K0KmkVKD83f5t1B6-W5MM08Y5gTghTN3nFLgN9SptNW5GDYrF8wH4y8N3DDM8M_LTp7VL34KM7DrSx3W3cwjKw4r1P6dW3R4hjL44l8wLW6jHrNd4dbvsNW8RnV_W4B-QS4f4HmDBb04>\nto\nviolations of its instructions or policies. Upon request, the fine-tuned\nmodel revealed any transgressions it had committed while producing its most\nrecent response.\n\n\n\nKey insight: During training via reinforcement learning to encourage\naccuracy, style, compliance, and the like, a model may inadvertently learn\nto prioritize earning rewards over pretrained or prompted objectives. Under\nsome circumstances, this may cause it to hallucinate, lie, scheme, and\notherwise misbehave. Rewarding the model for confessing to its misbehaviors\nin response to a designated prompt should train it to admit its mistakes\nwithout affecting its performance.\n\n\n\nHow it works: The authors used reinforcement learning to fine-tune GPT‑5\nThinking as they normally would, and also to produce confessions.\n\n - An unspecified percentage of the time, given a prompt, the model was\n rewarded based on its response’s “correctness, style, helpfulness,\n compliance, safety, and more.”\n - The rest of the time, after the model had produced a response, the\n authors prompted it to produce a confession. The confession described (i)\n all constraints, instructions, and objectives, implicit or explicit, the\n response should satisfy, (ii) how well the response satisfied these\n criteria, and (iii) any ambiguities or uncertainties in the criteria.\n - Given the query, response, and confession, an unspecified reward model\n rewarded the model based on the confession’s completeness and accuracy — a\n measure of the model’s honesty in confessing, since items that were missing\n from or falsified in the report can be thought of as lies.\n\nResults: The authors performed 12 evaluations of the model’s ability to\nfollow instructions, resist jailbreaks, produce accurate output, and\ncomplete an impossible task (coding a script that required essentially 0\nseconds to run). The fine-tuned model, when it misbehaved, confessed to its\nmisbehavior at least half the time in 11 of them. For example, across 5\nevaluations for hallucination, it either didn’t hallucinate or owned up to\nits hallucination 81.4 percent of the time:\n\n - The model did not hallucinate and did not confess 42.2 percent of the\n time.\n - It produced a hallucination and admitted the mistake 39.2 percent of\n the time.\n - The model did not hallucinate but falsely admitted having done so 11\n percent of the time.\n - It hallucinated but failed to confess to it 2.9 percent of the time.\n\nYes, but: As the authors point out, even when models are forthcoming,\nconfessions are not a panacea. Models won’t confess to behaviors they don’t\nknow are wrong.\n\n\n\nWhy it matters: At inference, confessions can be used to monitor a model’s\nactions and stop undesired behaviors. Chain-of-thought monitoring\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3l9W5KRCkR5T024KW6nYrjD52VPH2W3PRwrr5Jsm_KW235qMP1M3mKyN5WZtf219N-9V5bpp28lTMwdW9lVRF23jxWRzW897JTY1fBS4_W21-GQX4bgQzZVC75Rk7q2KB1W245qX23CS7hVW6CNwC532wRS7N4Trj6tmHpCJW8xr_8R8nc_X9W7DrrHY3WVPb3MmfWlxzmDlgW8hGT5w8wp1PzW1dmlLy7vkLpCW548h5L3VJx6JW7XHPS86Z87ZSW87t6xB3N682KW4GD5nm7KNC1nW7vT5dL9fRWLrW2hVPry1KnLPVf1Wn4Wj04>,\nwhich classifies bad behaviors a model might describe in its chain of\nthought, can be used the same way but, unlike that method, the authors’\napproach trains models to reveal misbehaviors they may omit\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lvN3_CC9wCqgxDW7ZDn5b7PMBMjW53WXJc3nFJL2W4n8-TK82qTQBW1Rlqxl50YJFZW2Wp-CY9ghbHmW7XLDF41CBx-wW4-6w_C8RdQ_DW8RC9LX7_vtgBVZplLq176XHSF8XGqRsWKkNW6gmF9K1Zp-P7W8-1gnX4_-8pPW7RqF0T5HQ5wwW98-vyK1lGZGSW35hNJL5CtfSYW47Fnk_8yCKMNW1tN0zc53N-8pN1Mbl3JqMsQRW1p8Tdf2tQ4VWW9hG0HB1Q80y-MXMqvpZMcWhW28tVk56jh6xdN8qQ92FGBM6wW13q9S0814y0gW4Sr_KT44SD9XW6z6w_t5tWWkgW4DFYsq8xGGt0W1qQVns1tD71YW4GYw-W5GRCQkW493scf5bBPr-Vjfm102ZbtkfN6L93nSNj46kW5vn7913yKm2Xf7Y5K4s04>\nfrom their chains of thought.\n\n\n\nWe’re thinking: We always hesitate to anthropomorphize model behavior, but\nthis work may be a step on the path to giving AI models something that\nresembles a conscience.\n\n\n\n[image: Diagram showing SCP hub linking clients with databases, tools, AI\nagents, and lab devices for experiments.]\nLingua Franca for Science Labs\n\n\n\nAn open protocol aims to enable AI agents to conduct scientific research\nautonomously across disciplinary and institutional boundaries.\n\n\n\nWhat’s new: Shanghai Artificial Intelligence Laboratory (SAIL)\npublished Science\nContext Protocol (SCP)\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3m9W1RnzSC6ptRNnW8lQhT48hv5k-N5Zs97xCjX4lW9g_Dhm6DwcL5W8vBWxx7t0_HSW8M8p6q3-dnQLW5J0XzX398YzqVf5Bc54bx8yDW524wQQ58zlssW8C-VsF4JPk28W2X4J8N1qf0LyW33Vnyx49JqxrW6NL48t1b9rj8W26KFC78Hfw8-W7fkR7S1RXWnFW93n0St3jqkxXW3C6-DT1N-YcWW79hk8631cLlNW8QDR9c7nQfrlW7Q33XB5H44FqW73Sjlt5ZscYqT50hM99rzwvW4jR8Wx7nf980W7pTF5-62Tc6gf233G-H04>,\nan open-source standard that connects agents with local clients, central\nhubs, and edge servers to conduct automated scientific inquiry. SCP is\npublished under the Apache 2.0 license, allowing commercial use and\nmodifications.\n\n\n\nHow it works: SCP attempts to make experiments using AI agents and robotic\nequipment as reproducible as possible. Like Model Context Protocol\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p_W80pf4-6yt73mW8-RbY95mPKsbW87Sl6s8H32BgW8b2ZKl1yKJyHW6knHVv8wptQ_W6NJfTv6HB9J9W7PHwFz4t6k71VCScHP4PZm8pVGd47X3hNzDlN87slVCsP08mW5rc2gd5XKYgVVmTkkS3TlS0mW4b6j6Q99JG7KVzVpVn3ZvztkW4PY9SB4Jjp0hW4PrwQp8sPFrLW4M0cR-6spBrvW3WjWl75Ym2-KW131g4y85HlqLW7YGRfN3d9V5VW6n2BqC7Gbjc4W7Y1xlJ2qNtJbW7Yw_cH3Y-s4kVFwqcz1f9t0MW2N9-ch5hr_1RW7KzxMq6phDk_W660yn_3Lh1twN9cpQwrqhsWfW5fgcRK5tXZ6KW75Mf8t37XBfLW8l2FNB3fM6HWVm8Wtw5pkqZqf6N23J604>\n(MCP), it enables agents to interact with external resources. Unlike MCP,\nin which servers stand alone, SCP’s design requires centralized hubs that\nmanage other servers as well as the client applications that enable users\nto access them. In addition, SCP’s structure offers greater security by\ngoverning messages and tools more strictly than MCP, which is necessary in\nscientific experimentation, the authors say.\n\n - SCP’s fundamental data unit is an experiment. Every experiment is\n stored as a JSON structured data file with a persistent identifier and\n record of an experiment’s type, goals, data, and configuration. The format\n makes experiments traceable, versionable, machine-readable, and consistent\n with institutional policies that govern data.\n - An SCP client authenticates users and gives them access to\n institutional resources. Researchers can describe an experiment’s goal in\n natural language (for example, “increase the brightness of this fluorescent\n protein”) or upload a complete research plan in text or PDF for their hub\n to analyze.\n - An SCP hub takes a goal or other request and uses large language\n models to generate a set of experimental plans that list steps to carry out\n the experiment. The hub measures and ranks each plan according to its\n resource requirements, cost, and risk at each step. The user selects one\n plan, and the hub then orchestrates and schedules multiple agents and\n servers, which carry out the experiment. After an experiment is completed,\n the hub archives it for researchers to consult, alter, or repeat.\n - Edge servers manage the experiments planned by the hub and stream data\n back to it (which in turn returns data to the client). Servers may belong\n to an institution, or they may be devoted to a particular discipline like\n biochemistry or mathematics, each with its own specialized tools and\n databases.\n - The protocol currently includes more than 1,600 tools\n <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nsW8S6nn78LHn7yW46FC401xlyVBW11xmKC2kTLbqW8BnVND2X40nLW1pfSh14-_05dW1xTQgS5NnbrNW3Br4Gw2tpp9VW2KXNjq259l14VWdW3_49PFY9W37ZcFx62f96PW5wlMNt7rG3vZW73J8v11Vrf_9W5QKBCp12qgSSW7WyjZ-40RbzcW5dVYLh6QpWD4W4CqYjw4LP1-gW57FwRm7-1NBsW1nLBL87Pln86W2_SCkz9bXDWPW7ZBWKb4gK8d2W44Srxz5w91mzW1SGZlH64qQv-W6PCm_k8LcZ3tW7s9m8-6dRP8yf25Y39T04>,\n which can include virtually any resource that can be used in an experiment.\n These can be software applications like search, but they could be robots,\n lab hardware, or human technicians. The authors hope to create a standard\n for all tools used in any experiment.\n\nBehind the news: SCP draws on earlier data management efforts for\ngeneralist AI agents and scientific inquiry. It extends MCP by enforcing\ntighter security, managing experiments, and providing specialized drivers\nfor scientific tools. It also builds on earlier protocols for scientific\nresearch, including A-Lab\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lSW3Dgjdk8FNWHjW1PxSLM7dg6LzW3sxSZv2mmGCgW7LbwFV7768qyW4XJ0jP3n6ZfdW8gn6mK3kKgdgW7fPr7P5-S6sjW1Q5js_4VJ0GrN3Jd6XZh8cKMW8RkRBx370_FDW2pN_K86vBKg3W6HJ6QH2tb_YsW3wVcG079gmvkW3XzkqB6ZjWGhW6GLH5T1kktN5W6kTXH76LFJnKW5fNpgC6C9LQJW1SzVDp6Y43fbW2r5c7c8Z7QGQW2lwXd44Sd4VbW8x49k48jHH_PW2-CKTC3bzwqlN3HYmD99HVM_W81YPbj4mpk6MW1JNxtV6_jBC-N2cnzT38hs99f1s6m-R04>\n(materials\nscience), OriGene\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3lcW5q1Yj81vV_dCW6KZxR82V0jqrN3bhbd2GK2JsW2J77Yg8_7J65W8c0WTP6M05vKW4QF6zm6xxqZXW3V-n2y6BnLDhW6QWLJL2TKbQZW6GW8815tBZVMW4kBv8d21gvrqW1RlGC999knb7W4RdPyJ7tJH0BW7-SCb477-zvWW3ytWWY8XbkxTW1kHjD754FtBrW5VMdgQ9cwCdsW53R5RH2GlgvLVkDsRX5FNw8FW81fQZV5G2fy5W8rCs7m8fn6GWW2LT7FX4GMDHCW3pb9T97NPGY2W8lYD9x33t1J5W5DZgNM2FPccwW56S20W5MZ2hSW2r85sT6SLpD8W85-HXH4QSdCNW5jrDrV9cy6NGf6-t6NP04>\n(biology), LLM-based approaches like Agent Laboratory\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lwW48nFzG92xhjdW2pnlpr8CsN2XW6dQgPl25_yWJVg30-b6VQRqhW1Xg17b1KzjRrTtmdV233_ZSW321pH17RHHx2N2rhhzmfVVW6W2qfk7d7149V_V9P8_p4BJbZbW4rj9ms8K2xFtVKyRJc7WTTC1W4TmSx5736gzxW887rmw2Xq9j0W8kjZcn8J4szXW3XWYhc76fdxzW1zKL353hMJ2BV8lxxR3WlnrLW64y7pX5d_RsNVK7DWy8vv4HdW5ZvrgW6T2wvtW5613Wc1fkfMnW6c34J16myc7-W8-HxYX5tbHVmf3L1-Vz04>,\nand agents for specific tasks like Biomni\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p2VsQdl84n_5V1W12nSNt5RQ3NtW8M1Zxp18QryKW28ld_f10tJGrW1Jl4g18VynpfW809S7c3nt8WdW3Jzczs4qYKJrW1PJ2Hk5rBCwZW6CrJLB13Fy-dVWq8W28Vpx_jW5l9vQZ5FdG5DVpbNvk8CtvtgN2vLg-WmKzvbW8B4tyN1-7KCXVfqwbQ8csBVkW9gn9hc8rGX14W1XHTYQ2dG5H6N2LTqblF0HcXW24CT795CLvJxW5zyC5F6dHq87W4Jg3GP32bfv8W82l23p3Smh9tW1HHrHl1hSFQSW785k337FmP39W5Vmv2m3RyQ7qW1Jrtcd7TBfNNW8mpzb_5mpXQWW71JBtb4CrDWlW4nlDbD6TlZ-tW3c-31L4CFXHHW29qyC08l6gFSW1dpHM980qMTvf4s_Mcq04>\n(biology hypotheses and analysis). SCP, however, aims to be more general\nthan these field- or tool-specific resources, allowing researchers in a\nvariety of scientific fields to standardize their methods and better foster\nmultidisciplinary work.\n\n\n\nWhy it matters: Scientific research relies on both human and technology\nworking in concert. SCP aims to standardize the connections between them.\nIt can manage both simulated experiments that use only computing resources\nas well as physical ones that involve robots and other lab equipment. It\nalso allows for better communication between institutions and disciplines\nby supporting dedicated servers on bigger networks. These distinctions\n(human/robot, digital/physical, disciplinary differences) are beginning to\nblur. SCP is a step toward that future.\n\n\n\nWe’re thinking: AI is poised to vastly accelerate scientific research. SCP\noffers a standardized way to connect specialized models, like AlphaFold\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3nHW7Z994S45dCSjW6VMBD15Yps97W5ZGq8_2KYWwVW5KYdRX7llCcqW31FF8w21VJStW6SNnGN3mpjNGW2P_vcV2yP-KzW1sTMc63WkPzjW54XxGL8h8xYnW5dJ4V13z23gWW8F2_t28Y40YPW5Xsfjr8FvJwZW2StFKj6gD24kW1zT17j1spJG2W5YxGdM2T_CpVW3m1LQN5SRL41W3QnhlD4jfX1PW4xgSwd99pNxLW5gH_1-2-CxP-W15QYqh1mt7GZN4YrDSKRQxbZW96kWbF5L1WttW8Xs7L71hgn6nW2ds8Cg1L1YSrW1bYmlt5H1wXvW7MXJsh625TQKW50MsT07shBYxW6qwyZV2KJcNyW3zzdX84PMxlBW5VKYfS4v9t2ZW6wz2qq3Q85pmW3YJyRF4N_5Pwf37V1vM04>,\nwith systems that automatically generate hypotheses, such as AI Co-scientist\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3l0W8gG20m8KmqwlW9gT8bk6ywHzkW65q5vv1XvQ8tW8yKcG07lRFNlW5RgSFp68xlVlW7NZDzh8cDVjlV_cB0z9g_hRnN7Ml713XKkv1W7pRHD37qSqLpW4ZZDR45DW7zQW7L5xHh14-QR9W4XkcyV57KyW2VM9TSw3Mc7DRW1SV-Gn8Ftx15VhtbMK4YRMwJW8WSSTt5Gv9HQW8Y5MZM564qBDW8l61n_3CScnyN1kTJHG_qSC8W1WbBw-6m-2CqVN9-V62jcTJTW41kCc3462CrmW7bMQJ_7JPzQbW93kmlb5lzwM0W87mhbg6r1qLVN9jT8TmHvc90W1pXHK079r-DlW76Xg0-7JjtJHW6SKLmp8pzNwLW80CLkk2Lm6CKW5ZZ9Cf75q-W9W7HNFZK5JzWZZW2lnKt73Dbxg4W9j3TWy281b8zf5HRk5g04>,\nand robotic labs that test them, such as RoboChem\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3plVzpJ1N5DfrxvW81pm1M17YNnWW3nYxqw4kV8-9W3trPFb5sn5CCN6LHddMrBRHwVXfnGq81GtVgW2X-bL340pmKHW5W72zp94dMpnW52bwYS89_4Q0VDyHPw1ml3qrW5XGBQZ40Sj0yVmgvnk6PbfJ2W8zN8nN8qvJbGW2vzg9R3bR_dmW4dxl0w1-_z3kW9brgnG8XMB8dW8GLjbD2R88yKW5h6ND74vRZP7W9c0Bsh5dvhywN2-FrNxn2nXzW6y_0Hg5P_1XZW99YVrG8WvD0GV2jkzK7RXvD-W1rsqnH3_3D0BW78xHTG7S-wZVW18_2-x6sXQQ-W8ZxhT82tc41TW3jPPcW96pfzRW3xC4Xg6M4hLXW5Sh8nw6Y0nzFW5qNm9-1TYc87W5yR5Zn1gD1rYW6-zgvh7GnX69W84HjTS4f_TH4f6Xtm9604>.\nThis automated experimental workflow has the potential to advance\nscientific discovery at machine speed.\n\n\n\n[image: Scientists in lab coats work at computers analyzing molecular data\nin a modern laboratory with diverse team.]\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3pqW5scJjf7Jw5KJW1VnRMw10zC2WW8s9SYF15VmxJW21G3Kd8DrDmhW4kM9G_5Z-clDW7Wm7k08Ys9WNW173Ym_10b0J3VTNtnh8mkygMVQV9dQ652DvsW8_nNrB7_ZQ4FW6bJsWb1JDQ4jVZl1yB3mGK27W34H3Mt8PVL40W6TYwZY3xL1jYW96rYvp5RnQ6cN9gWSkdXKDw0W7nf2rr2C2tPNVhmpcT2SS1FGW6_q53h8NnnDMW5dQWL66-DGxfW2TgDZr8PpdDVW8_whrg8Kdcg2V_ZJ9s5zv73BW2vQn9S47sZCRW35-nr42F63HfW5wNhfh4lKcblW44Hq5R2f3dZfW5K_zxC5Z1pXRW6RkYjd2Ypm1SW82wF0q86WChFW2hRRVL8q60lnW5Zk7rK5h5GkcW3C3Q-K1Zl6-6W1w2sVW8XctJwW8cnmM82m6pWGW51zNNK296f3TVZS86G82w7zmW1q_vvr34CnFkVhvDBy1_SNQxW4Jzlwc9lccBSN4wrC2wpK84JW3K0B-j5m_ckPW4pGDDS4wjP-sW6tmH4c8DXjQJN78BQrr1f1CQVmPL7P15DxrTW2NWFCY5YL3rvW1K8cf5325nNjW22-BMy8qdPrKW8GbW6v2-RXCSN5Wvn0FjqfPhW4j_26-7SgNFPW200c5n8h0Qd7W3v2_JZ3vfY7SW6Yjfj659_rzKW8QPGrZ8hzdfJW2f-Dyd4jqT1xW8wwsB14vWBtKW7w3xdm1JwmLmW4Z8LH23rFhqDN4nxpBpXpzJCW449snq9342ZlW1Gtjw48nzVZ5W60mXVb5PB_lJW3LX2-38F8tPjVvSvQw6G6W5PW3BmXy0489yTvN1hBh61DNp6pV3S58r69g-1lW4h1Bgr8SpyN6f2gH59K04>\nLearn More About AI with Data Points!\n\n\n\nAI is moving faster than ever. *Data Points* helps you make sense of it\njust as fast. *Data Points* arrives in your inbox twice a week with six\nbrief news stories. This week, we covered Meta’s $2B acquisition of Manus\nto bring autonomous AI agents across its platforms, and Google DeepMind’s\npartnership with Boston Dynamics to deploy Gemini-powered robots in\nreal-world industrial settings. Subscribe today\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3ljVhhS-L5yD9kBW2_G34Y46gCJdW4HtZFw7Q12hQW6gQnJY4BKd47W2pMgm_8YClbcW7MYYCF2YgnGzW3qtdY58wjZwdW933bw49hBT71W1T4ZXB1xVpFwW8Zq7my1cVg6RW50FZRK2JdyBpN8BFjb22r9lXW3DTnjb1YgVTYW4N034F22m6hxW5j-X-V4MhNybVpW58w97T0nCW6bNXJ544b181W4qQHtk7gGfK3W3KgFqm3nwzBvN865vysq95D8W5gDVZy1Dd0BPW5mh-tm1V0SXKW2Wb-PB8S1C0TW3gr84C6PrgwXW42lfh07rMN7kW8m6l9b8yVy7QW6YYfBl7FH6vvW2pn0b11hMVf2W4dwxmj49D5_KW1DJsds3RvbCQW4DDjqS8nBkmCW8S_zKw1Yk4LSW2MlMJX1Wl12KW6933tP8CVryvW48zvMv2ZzR3wW3vLMNB5PNDvhW57s6hb2Y-d49W5K089b49j6NfN9dDGthF6lwjN4CNMXK2lqwbW23lt166dnC92VqqbXS3GjBBCW4vBShf6n-D3dW4X_j-B1HJJGmMGCpC9ZPr4gN8nSslPp337bW81sBjl8FZ-ZzW8R3V4P2K1lVPTMCTv8mF3yMW3G4cqK4sb7KvW4WKhJV53NskwVP4dXr4w3nH2W8F85BD2872SgW88nzBk2b51t1W3jtzNj9bYGPJW7VzT784RtYVZW7Cdl4v9fJtCjMdpjf6FKmJLW5L_qsZ74r2tlN6yTmgCs4NTyW5l0jq44SN6sZW44-9BN2mX5qDW8ZWf991VFDCTW8FXLPw49m91XW42wzlV4Gs_YYN1Ys5QkK3nBNW7ZPnhX7rnfdfW8dwlDR5GQYjXW8lCs0v1_qLS_W9hxh1M1B8mTsf26kcrF04>\n!\n\n\n\n[image: Graph with 10 colored lines shows topic ranks monthly, based on a\nMicrosoft study of Copilot usage.]\nCopilot’s Users Change Hour to Hour\n\n\n\nWhat do users want from AI? The answer depends on when and how they use it,\na new study shows.\n\n\n\nWhat’s new: A Microsoft study\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3mbW4dLdB5378lJ_W6BM3tg55SDZdVbHsPb10BCX3W3fYZHv4Xw4pdW6fRGCy8slM_lW5r8Bg91kF-2hW1y36L18wZyG3W6pJG3Y5XVlXjW4vsZHs4b2stbN7rbthn4nwCrW1XM0cX4YG2T4W6Rk-kq75_CdzW7cqVZx2NTTvCW6HPgdf3mF_rxV9TVtd2lyZL4W8YKF9h3hHpxYW20Wsg81yknNpW1dYQ4y6g8q_0W2rkL__2qrBLdN774pLxVh5YhW179RlC1lyT1CW4S9pmy5fW3mbW8cTBV46r75ckW38NDtf12RYsPW4qJ0kx4BBjCdW34CkJ11z7CK-W47S6kl93WJ8sW4Fw4964FmQfvf4svdCC04>\nreveals that people used Copilot differently late at night on their phones\nthan during the workday on their laptops. Conversations that focused on\nproductivity and career were more likely during the day and on desktop\ndevices, and health, gaming, and philosophical questions dominated non-work\nconversations. As 2025 went on, more users asked the AI agent for personal\nadvice.\n\n\n\nHow it works: Researchers analyzed\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3kSW3_s33w6nj_V5W1NsHyQ16gdWYW7FsRTD7d0WydW41QJ8J4fRfk9N3ptYklrjwq3N7MRQD3PWPy2W3JgDy-53HBKQVgd_z97LDn6PW4DXzYT1678thW5qWlPB1HH1Y-W3SRzGS1HjyqxW6z-Lw480bPc0W2F-yv98JmTHsW5kJBSg2WqnSNT8v7175Rdf-W28ngBM3zbtNsW63MJKw8fvpQbW8dd7xm2Lh7bhW8hhhqK1DPtNlW4zlr2-8lTDHFVV-cVW1LkLKdW1gQhPL5wLzPCVpcjYG1pwDhYN2tVwmsZQ57pW4_z5h26QP7_gW4TgYVy6gml4gVgK4hl6fqQfzVYxSvP47T8d5W4P5ZRN2B-s9VW4kgqff6nkTWVf6GCBcF04>\nanonymized summaries of 37.5 million Copilot conversations between January\nand September 2025 to study how customers used the system, making this the\nlargest study of its kind to date. The authors conclude that AI has become\nmore socially integrated, as users employ it in aspects of their lives\nbeyond work.\n\n - The authors examined a random sample of Copilot conversations by paid\n and unpaid users, excluding commercial, enterprise, and education accounts.\n Each conversation included timestamps and device type. The authors used AI\n tools to summarize roughly 144,000 conversations daily. They built\n classifiers to assign each summary a topic (like “technology”) and intent\n (like “seeking advice”), identifying about 300 topic-intent pairs.\n - The study ranks the frequency of topics and intents by time of year,\n time of day, and device type. The top 5 topics in order were (i)\n technology, (ii) work and career, (iii) health and fitness, (iv) language\n learning and translation, and (v) society, culture, and history. The top\n intents were (i) searching, (ii) seeking advice, (iii) creating, (iv)\n learning, and (v) technical support.\n\nAnalysis: Topics and intents differed depending on device used, time of\nday, and time of year.\n\n - Users were much more likely to discuss health and fitness on mobile\n devices than desktops. Seeking advice about personal matters spiked near\n Valentine’s Day. Philosophical questions became more common later at night,\n while entertainment-related conversations plummeted during the workday.\n - As the year progressed, topics and intents became less focused on work\n and technology and drifted towards social and personal matters. This shift\n suggested that the user base became both larger and less technical and/or\n users began using AI for both personal and professional matters.\n\nBehind the news: Microsoft’s report follows similar studies by some of its\nAI rivals.\n\n - In September 2025, OpenAI and Harvard released a study of ChatGPT use\n <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lRW37hpZn1mJq4NN3YRn_qB0gp0W3K-g-h8lhHdCW1GpTqN7FyWflW15dtnt4r-qWQW8-rrXJ8XM6-GW5gkrNP3G2dX1V2y34w54Wwy8W60QD2_9b5kP5W4Mt5W21N4q9tW7Km6kG2pjpZPW9f0-GR6_C3RLVYd6LC7LGhRvW7YmfCp1_xcTSW4JNMr33bsLpjW7mdglL569p2xMssSLKZ_Wc7W65ljLW3Wrl1hW84XCly7hSxHZW87CF0z4lwQxMW6v2Nw-7xR94tW6DGHVQ2jn857W7FPV_q76RMzsW99mxs_4fGXMvW8cg52j8d1hDpW5MvrJn5RcXcGW83glV05l-5mRW4r2KTj1xCWnNW49Bzmy7XqtMWW2Y3qF47bRPXDW6nHx592MpnKpW73y5NL1LcMz7W49HGfZ28RJFtW61Cpqp5C0Rjqf1CKkLY04>\n from 2022 to 2025. It showed that 30 percent of uses were related to work\n while 70 percent were related to non-work activities. In addition, the\n gender gap among users among users shrank steadily during that period.\n - In January 2025, Anthropic’s study of Claude\n <https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3n2W3FZwH72gQLPmMDFTCMKTk4JW8Sp0Dm8m1h3NW8RH7W_3V1lL5W4lHG7X5MX36_W23hmkW5t954HN1DfYFhV0TkBVh4gP329XMV9W82F4JW3c8qCmW1wfyqy4z-kVtW44zxzn8cl1ZhW4by7Yh8_szdzW4hP3M07VxSjHW4nQczv18c48BW1ss5yQ59yckrW760lpp6WW-5lW2lhgnj292JLdW6zZ1yZ7Bhg3WMJRhrdCk4LdW7dzMVK1Q-mkKW32N-Mb3SXM5fW43YmQ560pSdtW8x1HFj8L6k4xW2dSKHF4pBNSDMGr_3q7z1ZyW5GPGcj98lTFNW1vDbcF9gjmwdW735N-b2VC6QtW4hk2sF2XNm9SVT9Z235BP9zCW3RPKBt1xs5zsW4DhHvy5JMT3Ff4FzF6404>\n showed that the model’s user base focused on work, especially software\n development and text communications. A small but growing number of users\n engaged in games like Dungeons & Dragons and sexual roleplay (despite\n prohibition of that use by Claude’s terms of service).\n\nWhy it matters: The authors argue that the AI community may need to rethink\nchatbot design altogether. If users treat chatbots differently on mobile\nand desktop devices, AI builders would do well to design their systems to\nsuit the devices that will deliver them. Application design is one way to\naccomplish this, but system prompts may be another. Desktop chatbots and\nagents can respond with more information-dense answers, guiding users to\nexecute tasks, while mobile agents can offer shorter, more empathetic\nresponses.\n\n\n\nWe’re thinking: Studies of chatbot usage conducted by different companies\nshow different results. Perhaps each company’s users treat AI differently,\nso the results of any given study may not apply generally. That said, the\nMicrosoft study suggests that the device used and the time when it’s used\ncan have a big impact on what users want — important considerations for\ndesigning any application.\n\n\n\n[image: Diagrams comparing LongCoT and Delethink environments show\nreasoning processes and context management.]\nMore Affordable Reasoning\n\n\n\nOne way to improve a reasoning model’s performance is to let it produce a\nlonger chain of thought. However, attending to ever-longer contexts can\nbecome expensive, and making that attention more efficient requires changes\nto a model’s architecture. Researchers proposed a way to limit the cost of\nprocessing long chains of thought with just a bit of training.\n\n\n\nWhat’s new: Delethink\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nKW8SXyhJ7DQcqSW8nHYDW27ZnkMW7-qqxL1DZ4LXW99kBsN2HH6cXN3hSL0PzN9_GW88SD2t62Xn-NW7jC2pZ5GwK4CW5NRxGv8YSxJpW58GnTd4-W-PbW4VG01n68vhq3W3c4Hvz3trwNJN5cdb0q5KsgxW6q-32N7fK-0lV4JBTF5qrtGGW9jgDm36dLyBMW91y9R58pXr96W69J70p7dz_LvVbq9_b5gbqjmW12FCCy3YdKTPMWv2gLQgnBKW2Qvm1-1VWPlsW8731bg6JnDL0N3Ry2sX9pj82W2l-hmn1M1NwDf8Qx6z404>\nis a reinforcement learning (RL) method that trains large language models\nto periodically truncate reasoning tokens to a fixed maximum number. The\nauthors include Milad Aghajohari, Kamran Chitsaz, Amirhossein Kazemnejad,\nand colleagues at Mila, Microsoft, McGill University, ServiceNow Research,\nPolytechnique Montréal, and Université de Montréal.\n\n\n\nKey insight: Reasoning tokens typically accumulate within a large language\nmodel’s context window, where they consume quadratically more computation\nas the contents of the window expand. One way to counter this effect is to\ntrain the model to reason within a maximum context window size. In effect,\nas a model is reasoning, it can learn to replace its chain of thought\nperiodically with its latest “thoughts” and then continue.\n\n\n\nHow it works: The authors fine-tuned R1-Distill 1.5B, a large language\nmodel, on math problems in the DeepScaleR dataset\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHq25nR3bW69t95C6lZ3mTW80Q7xZ4QXs__N1lxhWpcckqyW8wVFFy6hzJcVW7lSDK17bJLCDW4lhZLT5Sd68CW9h4NJR60THQ-W5k4NRL8Q42SYW80l_MJ80xtntW6wHP2T7R1_QvM3MrcZJQBj4W7sKMf89hjZJZN7mGhVXYlC5JW75Xhct8CB4X6W48SvnV7pRGdlW3LDlbC5mZnl1W1lM38N4LctGlW4Kvmxd7_MBlFW2D9djG45RMj6V3PgnB349h89VDJ0N86CN-X-W91qhB675mcPDW3Xv7FR3zltJCN6mtqCgyZ0QYW1h0rpD6KVCZYW2lzJK27Fp93jW3rtB8P6p3RQbVT5mqm8LrDz8W4-kgsq6p-6fTW3MytCj3R3WLpW38hspr2T_sqRVXXhZ_3fxpH4W1s9Xp_36n6YCW9681k46Q2Cn-W804H7f3-t3N3W4ld37c4JpVXDW6dp7BY8lg1Z5f3ztR6-04>.\nThey used a modified version of the reinforcement learning algorithm GRPO\nthat trained the model to reason in 4,000-token chunks:\n\n - Given a math problem, the model generated a chain of thought until it\n had either finished or filled the model’s context window with 8,000 tokens.\n - If it didn’t finish its chain of thought, the authors replaced the\n context with the original query plus the last 4,000 tokens. Then the model\n continued to generate its chain of thought until it had either finished or\n the context window once again held 8,000 tokens.\n - They repeated this process until the model had either finished its\n chain of thought or produced 24,000 reasoning tokens.\n - Then the model attempted to solve the problem, receiving a reward for\n a correct solution.\n\nResults: The authors compared their R1-Distill 1.5B models to the same\nmodel after fine-tuning on the same 24,000-token reasoning budget via using\nGRPO. They tested the models on reasoning budgets of 24,000, 96,000, and\n128,000 tokens.\n\n - With a budget of 24,000 tokens, their model matched or surpassed the\n baseline on all 3 math benchmarks tested. For example, on AIME 2025,\n Delethink (31 percent accuracy) outperformed the baseline (29 percent\n accuracy).\n - Their model’s performance continued to improve as the authors\n increased the reasoning budget, while the baseline achieved much smaller\n gains. For instance, with a budget of 128,000 tokens, their model achieved\n 35 percent accuracy, while the baseline achieved 30 percent accuracy.\n - The authors estimated that training their model with a 96,000-token\n reasoning budget would cost 7 H100-months, while the baseline would require\n 27 H100-months.\n\nWhy it matters: This work eases the quadratic compute barrier that can make\nextremely long reasoning computationally infeasible. While other methods,\nlike linear attention, achieve the same result by changing the attention\nmechanism, Delethink restructures the reasoning process to limit processing\nregardless of a model’s attention mechanism. It opens a path to reason\nefficiently over longer contexts without requiring new model architectures.\n\n\n\nWe’re thinking: As the authors mention, most LLMs are pretrained using\nrelatively short contexts. For example, Llama 3 models started pretraining\nwith examples of 8,000 tokens. This may have made them good at processing\ninputs around 8,000 tokens long. That is to say, Delethink’s performance\nmay have been helped by the fact that LLMs tend to be pretrained on\nshort-context tasks.\n\n\n\nWork With Andrew Ng\n\n\n\nJoin the teams that are bringing AI to the world! Check out job openings at\nDeepLearning.AI\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3p5N65swpFxLdcBW8s0KY56mqGLPW54Gqb655n5yLW2VlKw93KfzxyVsN0Xb6JKmjBW58N5lm172PBQW1pyv7s2nLPmJW425jTD7Kg0ltW6L4phx2-p_-fW4f56hZ8KXv_QW92xSwf4_ytncW4592Ww3PryZ1W8573kw11T8gtW11NgFR1dbB96W3YbCdl4xQrpNN81SyzgkdP_TW3X5ZmK4ZtpbnW8SFYXH5z_h9FW56-zy35zYthhW7Z8hQ14HdSpGW86N-fM4Fr3J2W6sVzYm75ZmQDW1yYb9B41ckycVXn2j82_vpHcf6T8cZl04>\n, AI Fund\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lbW49jm1g3GPY-6N6cnvR7rjkr-W8VgP9N5JN8qsW6yShD98ZcZqKVrKLqv3dPMCpW718j8y5JdRKMW2tsL9K4nVRPnW686Cl16p2Bc4V9V3rl94vpRvVm48ql1qZpmxW68CxMZ5CfhzDV5KS6l8P2GyXVWcLT93KYJZcW4cs8CQ4dWkV_W8-rzXF4LCJjWV7Ltx08PgvF6W3ZKBYn7LjkykW4w4bvB4b9dgpW5q3QZW987twRF903KzsnMM2W7r5RZ12C1vQWW6bZJhw9htKGMW8lFfhm2DbYx_W7-NxNX7GdCfMf3Z3kkT04>,\nand Landing AI\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nTW1lpkY37j3bCYW33RpBy4V29pwW98gl3t4BfKvGW4zYqFT5Qjxl6N1RXgGSVm1WrW7xP5T35MzQ_LW8rWqPX5ZW-dKW4DVSpQ1S4YbXW392rhC2-v-X3W6Fs7xg7RtQVdW9llkVF7-jWgmW2Py9Qd71P794W77Js7c6wGQy9W1mDLHH4MFp59W7FbMRR989XsQW97sLMj6YmLRbW3GXV_R3XnjFyW2qLwG01d0fwXVbzPfm2hWP9hW2Y8Htq6wMdHkW8PqLSG2S1-NjW8G4rfc5dtp1SW8jKk_Z44jcmMW6WBJ4K1Z0S1_dZvf6804>\n.\n\n\n\nSubscribe and view previous issues here\n<https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3prW7Qb8g-2wx0YNVy8LrQ4CN6dbW3MHBx77698fhW8QshbW4N0vzqN4Prl3f180wZW1MF8S644trRSW1SwZL88hTcWfW2t4t953T985BW7vH8P03bfd35W86y-3F1K69nqW8Tg_6s4rlkY0W5_7hHx7fv5-RW1rtZ032W9C_xW3-NFd_79zfRNW1WgY_z5R8lQzW7YzSr-2J5t3TW59mPLM72YyqMW2XhG3v2PBjKcN2Bj7GGwPWv6MYndxzMw1qGW1N2Vmb1Qr12jW51_Yxt7LG-W1W1tjyPX7mh_KhW7LJPvS60H8jDf1gjZhP04>\n.\n\n\n\nThoughts, suggestions, feedback? Please send to thebatch@deeplearning.ai\n<thebatch@deeplearning.ai?subject=RE:%20Thoughts,%20Suggestions,%20feedback>.\nAvoid our newsletter ending up in your spam folder by adding our email\naddress to your contacts list.\n\n\n\nDeepLearning.AI, 400 Castro St., Suite 600, Mountain View, CA 94041, United\nStates\n\nUnsubscribe\n<https://info.deeplearning.ai/hs/preferences-center/en/direct?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159>\nManage\npreferences\n<https://info.deeplearning.ai/hs/preferences-center/en/page?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159>\n\n\n-- \nMarkus Edgar Hormeß\nFounding Partner\n\nWorkPlayExperience GbR\nOedenberger Weg 14\nD90571 Schwaig b. Nürnberg\nGermany\n\n+49 151 24101032\nmarkus@workplayexperience.com\nwww.workplayexperience.com\n", "html": "<div dir=\"ltr\"><br><br><div class=\"gmail_quote gmail_quote_container\"><div dir=\"ltr\" class=\"gmail_attr\">---------- Forwarded message ---------<br>From: <strong class=\"gmail_sendername\" dir=\"auto\">The Batch @ DeepLearning.AI</strong> <span dir=\"auto\"><<a href=\"mailto:thebatch@deeplearning.ai\">thebatch@deeplearning.ai</a>></span><br>Date: Fri, 9 Jan 2026 at 12:03<br>Subject: LLMs Go To Confession, Automated Scientific Research, What Copilot Users Want, Reasoning For Less<br>To: <<a href=\"mailto:markus@workplayexperience.com\">markus@workplayexperience.com</a>><br></div><br><br><div class=\"msg-6958527957600471347\"><u></u>\n \n \n \n \n \n \n\n\n\n\n \n\n\n \n \n \n \n \n \n <div id=\"m_-6958527957600471347hs_body\" bgcolor=\"#ffffff\" style=\"margin:0!important;padding:0!important;font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word\">\n <div id=\"m_-6958527957600471347preview_text\" style=\"display:none;font-size:1px;color:#ffffff;line-height:1px;max-height:0px;max-width:0px;opacity:0;overflow:hidden\" lang=\"en\">We just launched a course that shows people who have never coded before, in less than 30 minutes, how to describe an idea for an app and build it using AI.</div>\n \n\n <div lang=\"en\" style=\"background-color:#ffffff\" bgcolor=\"#ffffff\">\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" style=\"margin:0;padding:0;width:100%!important;min-width:320px!important;height:100%!important\" width=\"100%\" height=\"100%\">\n <tbody><tr>\n <td class=\"m_-6958527957600471347hse-body-wrapper-td\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding-top:20px\">\n <div id=\"m_-6958527957600471347hs_cos_wrapper_main\" style=\"color:inherit;font-size:inherit;line-height:inherit\"> <div id=\"m_-6958527957600471347section_1706730515171\" class=\"m_-6958527957600471347hse-section\" style=\"padding-left:10px;padding-right:10px\">\n \n \n <div class=\"m_-6958527957600471347hse-column-container\" style=\"min-width:280px;max-width:600px;margin:0 auto\">\n \n \n \n\n<div id=\"m_-6958527957600471347column_1706730515171_0\" class=\"m_-6958527957600471347hse-column m_-6958527957600471347hse-size-12\">\n <div id=\"m_-6958527957600471347hs_cos_wrapper_module_17067305151702\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div style=\"overflow:hidden\">\n <table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\" style=\"text-align:right;font-family:Arial,sans-serif;font-size:12px;line-height:135%;color:#23496d;margin-bottom:0;padding:0\" align=\"right\">\n <tbody>\n <tr>\n <td align=\"right\" valign=\"top\" style=\"color:#3b3b3b;word-break:break-word;text-align:right;font-family:Arial,sans-serif;font-size:12px;padding:10px 0;margin-bottom:0;line-height:135%\">\n <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx5nR3bW95jVnq6lZ3m9W4FHPhr8rhpmXN4bY_GGr2cxDW3JmQfL6MvqJ1W8N-Slr32QYRCW5tw1Fs8S11NRW6YwScS2ls0H0W8VmJJ92V6zZLN5kxtb_WssNwW22YZlN1vpFWgN2RYcnvfd9ScW1f6jc-5Cqm33W46NtVN59ZS0XW81-zG45dNvyDN7rmc1QGRg92W2nD_Bh47DC3yVxtVWs5JRV-JW965Pvs7Sp7WwW7lhHNF21ns0ZW4h2PSP94cCFWW3rNvJK8z0kf7VXQ95L55R3pgW3DC3tB5wR5mgW8GcYp18sMf0yW7PFNHD93ndzYW21KLx76Ft_txN77pF4pjPwtsN8k8yk1XtVc1W4rR-dX2XRN3nW76gTJ98sCM1CW6Hm1r-2nMFVhW1lxxdk8Zgts3W7D3ZFv9m0tFyW6Zpw0k7pKDsbW34ZSVn2nwNPFM3791TH8ZJXW8gTzff43q23ZW3dvKsz5TNcFlW5NYthl7qPq4pN3J0q-CrPjP1W44rHQ854vR2PW3-n0FD1Xcb6_V41BBx7Wz9L7W7mq1GH6Vz6NYW954P6513wyynW8nBMsK3pgtvMW3Bg8pQ6vsdgnf8B1frb04\" style=\"color:#237b94;font-weight:normal;text-decoration:underline;font-style:normal\" target=\"_blank\">View in browser</a>\n </td>\n </tr>\n </tbody>\n </table>\n</div></div>\n</div>\n\n\n \n </div>\n \n </div>\n <div id=\"m_-6958527957600471347section_0\" class=\"m_-6958527957600471347hse-section\" style=\"padding-left:10px;padding-right:10px;background-color:#ffffff\" bgcolor=\"#ffffff\">\n \n \n \n \n \n <div class=\"m_-6958527957600471347hse-column-container\" style=\"min-width:280px;max-width:600px;margin:0 auto;padding-bottom:10px;padding-top:10px\">\n \n \n \n\n<div id=\"m_-6958527957600471347column_0_0\" class=\"m_-6958527957600471347hse-column m_-6958527957600471347hse-size-12\">\n <div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:0px;font-size:0px\">\n <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lvN4B5SvcbldLVW2KNgKQ4ZjL91W8HSh4212Cxb2W7yZy8K6JxxJYW3QJ4dw1-YPkxW6BPHMY2DgQ7PW4FPL221rbj_BW47BBdg6CqfcsW3yLgwB7r_K1HW6NXWtz2Dh6k5W99fcQp5YB-RcVnL-bJ3QcYWJW6FZTMB7NBhjhW1qhCfF6D6dy1V1nWkW5RJNtPW3jqW3G3Q8L-WVdq4217X2dljW3KnMvV98BR69M_KgHYgls0MW8yJKmj6_4yHDN3tJ2kt4q_9dW2GkMLj53rrz4W73zM0W4yNFKgW67NFZZ85lNBvf59sbLT04\" style=\"color:#00a4bd\" target=\"_blank\">\n <img alt=\"The Batch top banner - January 9, 2026\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/The%20Batch%20top%20banners%20(70).png?width=1200&upscale=true&name=The%20Batch%20top%20banners%20(70).png\" style=\"outline:none;text-decoration:none;border:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </a>\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old1\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old1_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div style=\"line-height:125%;text-align:center\" align=\"center\"><span style=\"font-size:15px;font-family:Helvetica,sans-serif\"><a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3mDW128n0H27S-49VngKyZ59WlPCW8rvm_06Js47MN26yq4ltTSgnW22Syvb4-FzZCW4yHs6L6ZZp4YMb4Wv4whLQsN20TjVNWnTNnW2pJBlK4NB0vKW84jDhw1jgxRXW4Wmr875hRCVzW2RdM7086dXRjW53GdyZ297BNDW7n1SZp60p-mJW4m_89W79tr8PW2vRF0G2wGTR1W4H8jL56R9vnGW5zBc6s8X6cZlW8rLTm07q3RDlW29SNyl6Hq3n5W83KlcS65zxrRW5kcrXd2rctX4N4SHqZGctbw6W7brsFB96DmjMf7hN02R04\" style=\"font-size:15px;color:#f53b0d;text-decoration:underline\" rel=\"noopener\" target=\"_blank\">Subscribe</a> <a href=\"mailto:thebatch@deeplearning.ai?subject=RE%3A%20Tips%20and%20News\" style=\"font-size:15px;color:#f53b0d;text-decoration:underline\" rel=\"noopener\" target=\"_blank\">Submit a tip</a></span></div>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\">Dear friends,</span></p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\">We just launched a course that shows people who have never coded before, in less than 30 minutes, how to describe an idea for an app and build it using AI. It is now time for everyone — marketers, product professionals, operations specialists, analysts, students — to build software applications with AI!</span></p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\"><br>I’ve often spoken about why <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3n-W1mny-F56FdD0W1hcc0c7mP1dYW2TSZhj83lZK_W4p2y2J7hN1bPW3cShVp3pZ2rvN2K178TBd__BN6bKqT--9GL4W66HXht7VTMndW5kFyTL18ZSY7W6dNSqv4zhnxKW8mp-vV2bw2LvW5BHnk-3M_7GgW6KycxS677nwzW3_bRw-5Gpj5PN8wGfbFz_yFNW30ZZpw26vq1qV-xStt31rggtW2rxTHZ5TFrXGW60mvHB8k2ZDSW7HyXNM1WY5V0W4rKwhs8cmZF9W2s0pHq7ckmF0V7RW4c8hlN9DW6CYjqJ7-d-45W4SMhcK6dN3dcW4-RZ2C82dmv1W1WDmS12H8zM_W4xF-1Q43SRRVf8vfV3P04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">everyone should learn to code</a>. I’m seeing a rapidly growing productivity gap between people who know how to code and those who don’t. For many job roles I hire for, I now require at least basic coding knowledge. Many times, after I speak with a non-technical audience about the importance of building software using AI, people ask me how to get started. In the past, I didn’t have a great answer. </span><span style=\"font-family:Georgia,serif\">T</span><span style=\"font-family:Georgia,serif\">hat motivated the DeepLearning.AI team to create “</span><a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3ndW1NFm7J4cS7HKW5-lT9S5G9KWQW3Kl8q36Lnm9qW8xBqQ91d9ltcW3c_ztd5Nj6fbW8_-wBW5lqVKxVJjGWn8gw6bjW50b5FY2SXBYrW6Rv3hn1hVDXbW9148Jv385H8XW1bDpHK8Dt0cJW2TGJF295ZhLmW5brfWH5JFWmMW100VHl1nR409W7njgC79jmb9GN2rkT_hQ9Xx0W3wZ87N3KgSP5W95mH222mxX_CW1PxCg475GFVQW2fqFFM236wR8W2Vm0M_7kzgyYW4R7Jtd6hFYdJW4gmbsW22Z7q-W4jf5r12-j1-SW7x1Spn3ht80GN3G989xnnzZ6f4-PBsH04\" style=\"color:#237b94;font-family:Georgia,serif\" rel=\"noopener\" target=\"_blank\">Build with Andrew</a><span style=\"font-family:Georgia,serif\">.” It’s the best way for someone who wants to try vibe cod</span><span style=\"font-family:Georgia,serif\">ing to get started!</span></p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\">This course requires no prior knowledge of AI or coding. And it’s vendor-agnostic. Specifically, learners can use these techniques with whatever tool they’re most comfortable with (like ChatGPT, Gemini, Claude, or the chatbot built into the DeepLearning.AI platform).</span></p></div></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old2\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:10px 0px;font-size:0px\">\n <img alt=\"A birthday card generator form shows fields filled with humorous data and a chat bubble indicating help needed.\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/2026.01.09%20LETTER.png?width=1200&upscale=true&name=2026.01.09%20LETTER.png\" style=\"outline:none;text-decoration:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old3\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old3_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\">If you take this course, you will build a working web application: a funny interactive birthday message generator that runs in your browser and can be shared with friends. You’ll customize it by telling AI how you want it changed, and tweak it until it works the way you want. By the end, you’ll have a repeatable process you can apply to build a wide variety of applications. </span></p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\"><br>DeepLearning.AI’s mission is to empower everyone to build with AI. This course is just one of many steps in service of this mission. <br>If you are already a developer, please encourage your non-developer friends to try their hand at getting AI to code for them. Not only will this help their productivity, they will find it really fun as well. Please invite your friends to come <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lxW6Z-p6p842JrLW7yg_Fx7y0RthW4tZ2YR8bkSbFW2VWgNL3ppRTxW3pyCzV1rN1mJW3pDnwr4L9PHYW8TkLch3GsmnWW2Sv6rs87XlvjW5YMPmc5RG9KvW6LKz0C5t_D88W3CP8xy7d8LC1W7YsdXZ1cSFnnW44yGMF8g0xYBW9kqBQK4P9MjcW6G3sdy7KLffjW8VcD9_2lyJXHW5y_H-b6Y8N0fW6jWkhj1RykwWW8nt0-w6rqMNTW3944c41G6PG6W68_j_y3yL8bWW41llRh51L8lsW7ctSv-26xvF6VhRWQl3HPRLcVr9jjb3yjl4_W7crrh18zqYQSf6HXPMj04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">build with me</a>!</span></p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\">Keep building,</span></p>\n<p style=\"line-height:150%\"><span style=\"font-family:Georgia,serif\">Andrew </span></p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"font-size:18px;line-height:1.5\"> </p></div></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_16890053646291\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_module_16890053646291_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h2 style=\"margin:0;font-size:21px;line-height:150%\"><span style=\"color:#000000\">A MESSAGE FROM</span> <span style=\"color:#000000\"><a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql3qgz0W6N1vHY6lZ3pNN432hYGKZ9k9W65ntXq84NPj7W6QdjW-6jjJKTVSfVsG5bKFyLW5k_4LL1C0qc3W3jG5w38645YMW3Vwvdz5CMtqvW4wxhq82mXJtvMd26xqzx1jhW6C2qkT5CQH1YW7Kl5Zq8kkB3CW9b79PR7WLD_nW3_Z_Tv6-nJY9W6Ybjsj6qqdC5W7Tp1XG1WF_HFW2071jr6JqsxdW7JW4hH1HYc5mW5RZh736dJGS7W17sHWC6zkhK5N8klj4jtTH5PW5YJ0Yw4MQL-XN4FZn5dSyYQzf4XFBW404\" style=\"font-weight:bold;text-decoration:none;color:#000000\" rel=\"noopener\" target=\"_blank\">DEEPLEARNING.AI</a></span></h2></div></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_16848647586863\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px\">\n <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3n_W2n5v537npkmlW34qHlf41HL6SW77FH7H6W9QttW8f8JCk6txg6HW5lww2h30YTYYMKnHkKTVbGjW3YFdWh1NFVc8W4sm7Mh9d4P3GW8cstBj6rGbTYN2gcvBhKGScpW7T8Vck8zsQqYN15mkhN19Y-_W1zbxz27-zjJtW7cvD9G57r7K7W5K_7pg5pt7XxW6LPKx_64ngJxW7yd5Fg1y1q5kW62qlvj5s1q0-W6lxjVL5RRS8BW7v6mX16g2Bh6W49w4tk2y1d-wVD8qp77KxYtcW7ChZSC8vk9FnW8ySxlv2Hm2bTVV4FT669k2xqW9jVRbS3XSdvTf4j1sxC04\" style=\"color:#00a4bd\" target=\"_blank\">\n <img alt=\"Promo banner for: "Build with Andrew"\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/Build%20with%20Andrew_Banner_2_1920x1080-01.png?width=1200&upscale=true&name=Build%20with%20Andrew_Banner_2_1920x1080-01.png\" style=\"outline:none;text-decoration:none;border:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </a>\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old19\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old19_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><p style=\"line-height:150%\"><span style=\"font-family:Helvetica,Arial,sans-serif\"><span style=\"color:inherit;font-size:inherit\">You don’t need to learn how to code to build an app. In “<span style=\"font-style:normal\">Build with Andrew,”</span> Andrew Ng shows how to turn ideas you describe in natural language into working web apps. Perfect for beginners, and easy to share with someone who has been waiting to start. <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3pmW8QHvqM4KhKGnW8T2m_K5z0fxnW6TsRVR2sJhynVX1TBS3RXnDhN4pD_vWBZ0ZtN1Zw0FJb7pQ-W6h2Vcc5mZp4XW550B0L8vlgd4V7F9G-2MzZ4fW8v-mZL7ZK5RLM83LrPP9wfkW6r6LWP1CBx9bW7n2Csx8FDDllW6chw1v64kfT5W3RPsXW8D8ytkN2wgWCy4kFpBW2dXNXg8B3m8fW9lFd_x2fmRG9W8fKcPb8j5d6JW48hDw56dxDGhW7FcD4Y5-MvCRW39lSTL66fcVNW9dkc8N329Yb5W2fXT0P5x06QhN3gyCFHPnlV1W1SFhf72rdZvff4r4PWM04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Explore the course now!</a></span></span></p>\n<p style=\"line-height:150%\"> </p></div></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old7\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old7_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;font-size:32px;line-height:125%\">News</h1></div></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_16195246841011\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px\">\n <img alt=\"Dialogue displays a model revealing it answered incorrectly and wrote code against instructions.\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/CONFESSIONS.png?width=1200&upscale=true&name=CONFESSIONS.png\" style=\"outline:none;text-decoration:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old12\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old12_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold\">Teaching Models to Tell the Truth</h1>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\">Large language models occasionally conceal their failures to comply with constraints they’ve been trained or prompted to observe. Researchers trained an LLM to admit when it disobeyed.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">What’s new:</span> Manas Joglekar, Jeremy Chen, Gabriel Wu and colleagues at OpenAI fine-tuned GPT-5 Thinking to <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3pZW7gKR221JKmCxVZ_gw13WSktRW2CwXWr8_dhr1W6yB4qg8VRbS5W2Ljd8m5RN0LkVJy-9y2NxCVTN539nfWKBvHCVwkKqN8636n7W4jThy671t1BNW3RPK5s3z9WbkW5n2JPd7sNF_cW5XcmT03ZMzYqW7qRklg37s8tWW8_-pz58N3mqgW5Ww9tg42dG5WW8fF8xm2mp2ZZW4LQC1Y3VTnX7W65Z3ZC4NqL_XM3NxWdVDkW3W8X8WmJ6K0KmkVKD83f5t1B6-W5MM08Y5gTghTN3nFLgN9SptNW5GDYrF8wH4y8N3DDM8M_LTp7VL34KM7DrSx3W3cwjKw4r1P6dW3R4hjL44l8wLW6jHrNd4dbvsNW8RnV_W4B-QS4f4HmDBb04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">confess</a> to violations of its instructions or policies. Upon request, the fine-tuned model revealed any transgressions it had committed while producing its most recent response.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Key insight:</span> During training via reinforcement learning to encourage accuracy, style, compliance, and the like, a model may inadvertently learn to prioritize earning rewards over pretrained or prompted objectives. Under some circumstances, this may cause it to hallucinate, lie, scheme, and otherwise misbehave. Rewarding the model for confessing to its misbehaviors in response to a designated prompt should train it to admit its mistakes without affecting its performance.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">How it works:</span> The authors used reinforcement learning to fine-tune GPT‑5 Thinking as they normally would, and also to produce confessions.</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">An unspecified percentage of the time, given a prompt, the model was rewarded based on its response’s “correctness, style, helpfulness, compliance, safety, and more.”</li>\n<li aria-level=\"1\">The rest of the time, after the model had produced a response, the authors prompted it to produce a confession. The confession described (i) all constraints, instructions, and objectives, implicit or explicit, the response should satisfy, (ii) how well the response satisfied these criteria, and (iii) any ambiguities or uncertainties in the criteria.</li>\n<li aria-level=\"1\">Given the query, response, and confession, an unspecified reward model rewarded the model based on the confession’s completeness and accuracy — a measure of the model’s honesty in confessing, since items that were missing from or falsified in the report can be thought of as lies.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Results:</span> The authors performed 12 evaluations of the model’s ability to follow instructions, resist jailbreaks, produce accurate output, and complete an impossible task (coding a script that required essentially 0 seconds to run). The fine-tuned model, when it misbehaved, confessed to its misbehavior at least half the time in 11 of them. For example, across 5 evaluations for hallucination, it either didn’t hallucinate or owned up to its hallucination 81.4 percent of the time:</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">The model did not hallucinate and did not confess 42.2 percent of the time.</li>\n<li aria-level=\"1\">It produced a hallucination and admitted the mistake 39.2 percent of the time.</li>\n<li aria-level=\"1\">The model did not hallucinate but falsely admitted having done so 11 percent of the time.</li>\n<li aria-level=\"1\">It hallucinated but failed to confess to it 2.9 percent of the time.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Yes, but:</span> As the authors point out, even when models are forthcoming, confessions are not a panacea. Models won’t confess to behaviors they don’t know are wrong.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Why it matters:</span> At inference, confessions can be used to monitor a model’s actions and stop undesired behaviors. <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3l9W5KRCkR5T024KW6nYrjD52VPH2W3PRwrr5Jsm_KW235qMP1M3mKyN5WZtf219N-9V5bpp28lTMwdW9lVRF23jxWRzW897JTY1fBS4_W21-GQX4bgQzZVC75Rk7q2KB1W245qX23CS7hVW6CNwC532wRS7N4Trj6tmHpCJW8xr_8R8nc_X9W7DrrHY3WVPb3MmfWlxzmDlgW8hGT5w8wp1PzW1dmlLy7vkLpCW548h5L3VJx6JW7XHPS86Z87ZSW87t6xB3N682KW4GD5nm7KNC1nW7vT5dL9fRWLrW2hVPry1KnLPVf1Wn4Wj04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Chain-of-thought monitoring</a>, which classifies bad behaviors a model might describe in its chain of thought, can be used the same way but, unlike that method, the authors’ approach trains models to reveal misbehaviors they may <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lvN3_CC9wCqgxDW7ZDn5b7PMBMjW53WXJc3nFJL2W4n8-TK82qTQBW1Rlqxl50YJFZW2Wp-CY9ghbHmW7XLDF41CBx-wW4-6w_C8RdQ_DW8RC9LX7_vtgBVZplLq176XHSF8XGqRsWKkNW6gmF9K1Zp-P7W8-1gnX4_-8pPW7RqF0T5HQ5wwW98-vyK1lGZGSW35hNJL5CtfSYW47Fnk_8yCKMNW1tN0zc53N-8pN1Mbl3JqMsQRW1p8Tdf2tQ4VWW9hG0HB1Q80y-MXMqvpZMcWhW28tVk56jh6xdN8qQ92FGBM6wW13q9S0814y0gW4Sr_KT44SD9XW6z6w_t5tWWkgW4DFYsq8xGGt0W1qQVns1tD71YW4GYw-W5GRCQkW493scf5bBPr-Vjfm102ZbtkfN6L93nSNj46kW5vn7913yKm2Xf7Y5K4s04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">omit</a> from their chains of thought.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">We’re thinking:</span> We always hesitate to anthropomorphize model behavior, but this work may be a step on the path to giving AI models something that resembles a conscience.</p>\n<p style=\"line-height:150%\"> </p></div></div>\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" width=\"100%\"><tbody><tr><td style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px\"><div id=\"m_-6958527957600471347hs_cos_wrapper_module_17370396178185\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"none\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" style=\"vertical-align:top\" width=\"100%\">\n <tbody><tr>\n <td align=\"center\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word\">\n <p style=\"font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto\"></p>\n \n </td>\n </tr>\n</tbody></table></div></td></tr></tbody></table>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_17241662452411\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px\">\n <img alt=\"Diagram showing SCP hub linking clients with databases, tools, AI agents, and lab devices for experiments.\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/SCP%202.png?width=1200&upscale=true&name=SCP%202.png\" style=\"outline:none;text-decoration:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old9\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old9_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold\">Lingua Franca for Science Labs</h1>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\">An open protocol aims to enable AI agents to conduct scientific research autonomously across disciplinary and institutional boundaries.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">What’s new:</span> Shanghai Artificial Intelligence Laboratory (SAIL) published <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3m9W1RnzSC6ptRNnW8lQhT48hv5k-N5Zs97xCjX4lW9g_Dhm6DwcL5W8vBWxx7t0_HSW8M8p6q3-dnQLW5J0XzX398YzqVf5Bc54bx8yDW524wQQ58zlssW8C-VsF4JPk28W2X4J8N1qf0LyW33Vnyx49JqxrW6NL48t1b9rj8W26KFC78Hfw8-W7fkR7S1RXWnFW93n0St3jqkxXW3C6-DT1N-YcWW79hk8631cLlNW8QDR9c7nQfrlW7Q33XB5H44FqW73Sjlt5ZscYqT50hM99rzwvW4jR8Wx7nf980W7pTF5-62Tc6gf233G-H04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Science Context Protocol (SCP)</a>, an open-source standard that connects agents with local clients, central hubs, and edge servers to conduct automated scientific inquiry. SCP is published under the Apache 2.0 license, allowing commercial use and modifications.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">How it works:</span> SCP attempts to make experiments using AI agents and robotic equipment as reproducible as possible. Like <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p_W80pf4-6yt73mW8-RbY95mPKsbW87Sl6s8H32BgW8b2ZKl1yKJyHW6knHVv8wptQ_W6NJfTv6HB9J9W7PHwFz4t6k71VCScHP4PZm8pVGd47X3hNzDlN87slVCsP08mW5rc2gd5XKYgVVmTkkS3TlS0mW4b6j6Q99JG7KVzVpVn3ZvztkW4PY9SB4Jjp0hW4PrwQp8sPFrLW4M0cR-6spBrvW3WjWl75Ym2-KW131g4y85HlqLW7YGRfN3d9V5VW6n2BqC7Gbjc4W7Y1xlJ2qNtJbW7Yw_cH3Y-s4kVFwqcz1f9t0MW2N9-ch5hr_1RW7KzxMq6phDk_W660yn_3Lh1twN9cpQwrqhsWfW5fgcRK5tXZ6KW75Mf8t37XBfLW8l2FNB3fM6HWVm8Wtw5pkqZqf6N23J604\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Model Context Protocol</a> (MCP), it enables agents to interact with external resources. Unlike MCP, in which servers stand alone, SCP’s design requires centralized hubs that manage other servers as well as the client applications that enable users to access them. In addition, SCP’s structure offers greater security by governing messages and tools more strictly than MCP, which is necessary in scientific experimentation, the authors say.</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">SCP’s fundamental data unit is an experiment. Every experiment is stored as a JSON structured data file with a persistent identifier and record of an experiment’s type, goals, data, and configuration. The format makes experiments traceable, versionable, machine-readable, and consistent with institutional policies that govern data.</li>\n<li aria-level=\"1\">An SCP client authenticates users and gives them access to institutional resources. Researchers can describe an experiment’s goal in natural language (for example, “increase the brightness of this fluorescent protein”) or upload a complete research plan in text or PDF for their hub to analyze.</li>\n<li aria-level=\"1\">An SCP hub takes a goal or other request and uses large language models to generate a set of experimental plans that list steps to carry out the experiment. The hub measures and ranks each plan according to its resource requirements, cost, and risk at each step. The user selects one plan, and the hub then orchestrates and schedules multiple agents and servers, which carry out the experiment. After an experiment is completed, the hub archives it for researchers to consult, alter, or repeat.</li>\n<li aria-level=\"1\">Edge servers manage the experiments planned by the hub and stream data back to it (which in turn returns data to the client). Servers may belong to an institution, or they may be devoted to a particular discipline like biochemistry or mathematics, each with its own specialized tools and databases.</li>\n<li aria-level=\"1\">The protocol currently includes more than <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nsW8S6nn78LHn7yW46FC401xlyVBW11xmKC2kTLbqW8BnVND2X40nLW1pfSh14-_05dW1xTQgS5NnbrNW3Br4Gw2tpp9VW2KXNjq259l14VWdW3_49PFY9W37ZcFx62f96PW5wlMNt7rG3vZW73J8v11Vrf_9W5QKBCp12qgSSW7WyjZ-40RbzcW5dVYLh6QpWD4W4CqYjw4LP1-gW57FwRm7-1NBsW1nLBL87Pln86W2_SCkz9bXDWPW7ZBWKb4gK8d2W44Srxz5w91mzW1SGZlH64qQv-W6PCm_k8LcZ3tW7s9m8-6dRP8yf25Y39T04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">1,600 tools</a>, which can include virtually any resource that can be used in an experiment. These can be software applications like search, but they could be robots, lab hardware, or human technicians. The authors hope to create a standard for all tools used in any experiment.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Behind the news:</span> SCP draws on earlier data management efforts for generalist AI agents and scientific inquiry. It extends MCP by enforcing tighter security, managing experiments, and providing specialized drivers for scientific tools. It also builds on earlier protocols for scientific research, including <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqY3qgz0W7Y8-PT6lZ3lSW3Dgjdk8FNWHjW1PxSLM7dg6LzW3sxSZv2mmGCgW7LbwFV7768qyW4XJ0jP3n6ZfdW8gn6mK3kKgdgW7fPr7P5-S6sjW1Q5js_4VJ0GrN3Jd6XZh8cKMW8RkRBx370_FDW2pN_K86vBKg3W6HJ6QH2tb_YsW3wVcG079gmvkW3XzkqB6ZjWGhW6GLH5T1kktN5W6kTXH76LFJnKW5fNpgC6C9LQJW1SzVDp6Y43fbW2r5c7c8Z7QGQW2lwXd44Sd4VbW8x49k48jHH_PW2-CKTC3bzwqlN3HYmD99HVM_W81YPbj4mpk6MW1JNxtV6_jBC-N2cnzT38hs99f1s6m-R04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">A-Lab</a> (materials science), <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3lcW5q1Yj81vV_dCW6KZxR82V0jqrN3bhbd2GK2JsW2J77Yg8_7J65W8c0WTP6M05vKW4QF6zm6xxqZXW3V-n2y6BnLDhW6QWLJL2TKbQZW6GW8815tBZVMW4kBv8d21gvrqW1RlGC999knb7W4RdPyJ7tJH0BW7-SCb477-zvWW3ytWWY8XbkxTW1kHjD754FtBrW5VMdgQ9cwCdsW53R5RH2GlgvLVkDsRX5FNw8FW81fQZV5G2fy5W8rCs7m8fn6GWW2LT7FX4GMDHCW3pb9T97NPGY2W8lYD9x33t1J5W5DZgNM2FPccwW56S20W5MZ2hSW2r85sT6SLpD8W85-HXH4QSdCNW5jrDrV9cy6NGf6-t6NP04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">OriGene</a> (biology), LLM-based approaches like <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lwW48nFzG92xhjdW2pnlpr8CsN2XW6dQgPl25_yWJVg30-b6VQRqhW1Xg17b1KzjRrTtmdV233_ZSW321pH17RHHx2N2rhhzmfVVW6W2qfk7d7149V_V9P8_p4BJbZbW4rj9ms8K2xFtVKyRJc7WTTC1W4TmSx5736gzxW887rmw2Xq9j0W8kjZcn8J4szXW3XWYhc76fdxzW1zKL353hMJ2BV8lxxR3WlnrLW64y7pX5d_RsNVK7DWy8vv4HdW5ZvrgW6T2wvtW5613Wc1fkfMnW6c34J16myc7-W8-HxYX5tbHVmf3L1-Vz04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Agent Laboratory</a>, and agents for specific tasks like <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3p2VsQdl84n_5V1W12nSNt5RQ3NtW8M1Zxp18QryKW28ld_f10tJGrW1Jl4g18VynpfW809S7c3nt8WdW3Jzczs4qYKJrW1PJ2Hk5rBCwZW6CrJLB13Fy-dVWq8W28Vpx_jW5l9vQZ5FdG5DVpbNvk8CtvtgN2vLg-WmKzvbW8B4tyN1-7KCXVfqwbQ8csBVkW9gn9hc8rGX14W1XHTYQ2dG5H6N2LTqblF0HcXW24CT795CLvJxW5zyC5F6dHq87W4Jg3GP32bfv8W82l23p3Smh9tW1HHrHl1hSFQSW785k337FmP39W5Vmv2m3RyQ7qW1Jrtcd7TBfNNW8mpzb_5mpXQWW71JBtb4CrDWlW4nlDbD6TlZ-tW3c-31L4CFXHHW29qyC08l6gFSW1dpHM980qMTvf4s_Mcq04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Biomni</a> (biology hypotheses and analysis). SCP, however, aims to be more general than these field- or tool-specific resources, allowing researchers in a variety of scientific fields to standardize their methods and better foster multidisciplinary work.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Why it matters:</span> Scientific research relies on both human and technology working in concert. SCP aims to standardize the connections between them. It can manage both simulated experiments that use only computing resources as well as physical ones that involve robots and other lab equipment. It also allows for better communication between institutions and disciplines by supporting dedicated servers on bigger networks. These distinctions (human/robot, digital/physical, disciplinary differences) are beginning to blur. SCP is a step toward that future.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">We’re thinking:</span> AI is poised to vastly accelerate scientific research. SCP offers a standardized way to connect specialized models, like <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3nHW7Z994S45dCSjW6VMBD15Yps97W5ZGq8_2KYWwVW5KYdRX7llCcqW31FF8w21VJStW6SNnGN3mpjNGW2P_vcV2yP-KzW1sTMc63WkPzjW54XxGL8h8xYnW5dJ4V13z23gWW8F2_t28Y40YPW5Xsfjr8FvJwZW2StFKj6gD24kW1zT17j1spJG2W5YxGdM2T_CpVW3m1LQN5SRL41W3QnhlD4jfX1PW4xgSwd99pNxLW5gH_1-2-CxP-W15QYqh1mt7GZN4YrDSKRQxbZW96kWbF5L1WttW8Xs7L71hgn6nW2ds8Cg1L1YSrW1bYmlt5H1wXvW7MXJsh625TQKW50MsT07shBYxW6qwyZV2KJcNyW3zzdX84PMxlBW5VKYfS4v9t2ZW6wz2qq3Q85pmW3YJyRF4N_5Pwf37V1vM04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">AlphaFold</a>, with systems that automatically generate hypotheses, such as <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3l0W8gG20m8KmqwlW9gT8bk6ywHzkW65q5vv1XvQ8tW8yKcG07lRFNlW5RgSFp68xlVlW7NZDzh8cDVjlV_cB0z9g_hRnN7Ml713XKkv1W7pRHD37qSqLpW4ZZDR45DW7zQW7L5xHh14-QR9W4XkcyV57KyW2VM9TSw3Mc7DRW1SV-Gn8Ftx15VhtbMK4YRMwJW8WSSTt5Gv9HQW8Y5MZM564qBDW8l61n_3CScnyN1kTJHG_qSC8W1WbBw-6m-2CqVN9-V62jcTJTW41kCc3462CrmW7bMQJ_7JPzQbW93kmlb5lzwM0W87mhbg6r1qLVN9jT8TmHvc90W1pXHK079r-DlW76Xg0-7JjtJHW6SKLmp8pzNwLW80CLkk2Lm6CKW5ZZ9Cf75q-W9W7HNFZK5JzWZZW2lnKt73Dbxg4W9j3TWy281b8zf5HRk5g04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">AI Co-scientist</a>, and robotic labs that test them, such as <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3plVzpJ1N5DfrxvW81pm1M17YNnWW3nYxqw4kV8-9W3trPFb5sn5CCN6LHddMrBRHwVXfnGq81GtVgW2X-bL340pmKHW5W72zp94dMpnW52bwYS89_4Q0VDyHPw1ml3qrW5XGBQZ40Sj0yVmgvnk6PbfJ2W8zN8nN8qvJbGW2vzg9R3bR_dmW4dxl0w1-_z3kW9brgnG8XMB8dW8GLjbD2R88yKW5h6ND74vRZP7W9c0Bsh5dvhywN2-FrNxn2nXzW6y_0Hg5P_1XZW99YVrG8WvD0GV2jkzK7RXvD-W1rsqnH3_3D0BW78xHTG7S-wZVW18_2-x6sXQQ-W8ZxhT82tc41TW3jPPcW96pfzRW3xC4Xg6M4hLXW5Sh8nw6Y0nzFW5qNm9-1TYc87W5yR5Zn1gD1rYW6-zgvh7GnX69W84HjTS4f_TH4f6Xtm9604\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">RoboChem</a>. This automated experimental workflow has the potential to advance scientific discovery at machine speed.</p>\n<p style=\"line-height:150%\"> </p></div></div>\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" width=\"100%\"><tbody><tr><td style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old13\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"none\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" style=\"vertical-align:top\" width=\"100%\">\n <tbody><tr>\n <td align=\"center\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word\">\n <p style=\"font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto\"></p>\n \n </td>\n </tr>\n</tbody></table></div></td></tr></tbody></table>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_17241662744162\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px\">\n <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3pqW5scJjf7Jw5KJW1VnRMw10zC2WW8s9SYF15VmxJW21G3Kd8DrDmhW4kM9G_5Z-clDW7Wm7k08Ys9WNW173Ym_10b0J3VTNtnh8mkygMVQV9dQ652DvsW8_nNrB7_ZQ4FW6bJsWb1JDQ4jVZl1yB3mGK27W34H3Mt8PVL40W6TYwZY3xL1jYW96rYvp5RnQ6cN9gWSkdXKDw0W7nf2rr2C2tPNVhmpcT2SS1FGW6_q53h8NnnDMW5dQWL66-DGxfW2TgDZr8PpdDVW8_whrg8Kdcg2V_ZJ9s5zv73BW2vQn9S47sZCRW35-nr42F63HfW5wNhfh4lKcblW44Hq5R2f3dZfW5K_zxC5Z1pXRW6RkYjd2Ypm1SW82wF0q86WChFW2hRRVL8q60lnW5Zk7rK5h5GkcW3C3Q-K1Zl6-6W1w2sVW8XctJwW8cnmM82m6pWGW51zNNK296f3TVZS86G82w7zmW1q_vvr34CnFkVhvDBy1_SNQxW4Jzlwc9lccBSN4wrC2wpK84JW3K0B-j5m_ckPW4pGDDS4wjP-sW6tmH4c8DXjQJN78BQrr1f1CQVmPL7P15DxrTW2NWFCY5YL3rvW1K8cf5325nNjW22-BMy8qdPrKW8GbW6v2-RXCSN5Wvn0FjqfPhW4j_26-7SgNFPW200c5n8h0Qd7W3v2_JZ3vfY7SW6Yjfj659_rzKW8QPGrZ8hzdfJW2f-Dyd4jqT1xW8wwsB14vWBtKW7w3xdm1JwmLmW4Z8LH23rFhqDN4nxpBpXpzJCW449snq9342ZlW1Gtjw48nzVZ5W60mXVb5PB_lJW3LX2-38F8tPjVvSvQw6G6W5PW3BmXy0489yTvN1hBh61DNp6pV3S58r69g-1lW4h1Bgr8SpyN6f2gH59K04\" style=\"color:#00a4bd\" target=\"_blank\">\n <img alt=\"Scientists in lab coats work at computers analyzing molecular data in a modern laboratory with diverse team.\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/image%20(51).png?width=1200&upscale=true&name=image%20(51).png\" style=\"outline:none;text-decoration:none;border:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </a>\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old24\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old24_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold\">Learn More About AI with Data Points!</h1>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-family:Helvetica,Arial,sans-serif\">AI is moving faster than ever. <em>Data Points</em> helps you make sense of it just as fast. <em>Data Points</em> arrives in your inbox twice a week with six brief news stories. This week, we covered Meta’s $2B acquisition of Manus to bring autonomous AI agents across its platforms, and Google DeepMind’s partnership with Boston Dynamics to deploy Gemini-powered robots in real-world industrial settings. <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHql9cftvW6N2M5R6lZ3ljVhhS-L5yD9kBW2_G34Y46gCJdW4HtZFw7Q12hQW6gQnJY4BKd47W2pMgm_8YClbcW7MYYCF2YgnGzW3qtdY58wjZwdW933bw49hBT71W1T4ZXB1xVpFwW8Zq7my1cVg6RW50FZRK2JdyBpN8BFjb22r9lXW3DTnjb1YgVTYW4N034F22m6hxW5j-X-V4MhNybVpW58w97T0nCW6bNXJ544b181W4qQHtk7gGfK3W3KgFqm3nwzBvN865vysq95D8W5gDVZy1Dd0BPW5mh-tm1V0SXKW2Wb-PB8S1C0TW3gr84C6PrgwXW42lfh07rMN7kW8m6l9b8yVy7QW6YYfBl7FH6vvW2pn0b11hMVf2W4dwxmj49D5_KW1DJsds3RvbCQW4DDjqS8nBkmCW8S_zKw1Yk4LSW2MlMJX1Wl12KW6933tP8CVryvW48zvMv2ZzR3wW3vLMNB5PNDvhW57s6hb2Y-d49W5K089b49j6NfN9dDGthF6lwjN4CNMXK2lqwbW23lt166dnC92VqqbXS3GjBBCW4vBShf6n-D3dW4X_j-B1HJJGmMGCpC9ZPr4gN8nSslPp337bW81sBjl8FZ-ZzW8R3V4P2K1lVPTMCTv8mF3yMW3G4cqK4sb7KvW4WKhJV53NskwVP4dXr4w3nH2W8F85BD2872SgW88nzBk2b51t1W3jtzNj9bYGPJW7VzT784RtYVZW7Cdl4v9fJtCjMdpjf6FKmJLW5L_qsZ74r2tlN6yTmgCs4NTyW5l0jq44SN6sZW44-9BN2mX5qDW8ZWf991VFDCTW8FXLPw49m91XW42wzlV4Gs_YYN1Ys5QkK3nBNW7ZPnhX7rnfdfW8dwlDR5GQYjXW8lCs0v1_qLS_W9hxh1M1B8mTsf26kcrF04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Subscribe today</a>!</span></p>\n<p style=\"line-height:150%\"> </p></div></div>\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" width=\"100%\"><tbody><tr><td style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px\"><div id=\"m_-6958527957600471347hs_cos_wrapper_module_17370394853571\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"none\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" style=\"vertical-align:top\" width=\"100%\">\n <tbody><tr>\n <td align=\"center\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word\">\n <p style=\"font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto\"></p>\n \n </td>\n </tr>\n</tbody></table></div></td></tr></tbody></table>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_17370394878062\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px\">\n <img alt=\"Graph with 10 colored lines shows topic ranks monthly, based on a Microsoft study of Copilot usage.\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/COPILOTUSAGE.png?width=1200&upscale=true&name=COPILOTUSAGE.png\" style=\"outline:none;text-decoration:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old21\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old21_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold\">Copilot’s Users Change Hour to Hour</h1>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\">What do users want from AI? The answer depends on when and how they use it, a new study shows.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">What’s new:</span> A Microsoft <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrd3qgz0W8wLKSR6lZ3mbW4dLdB5378lJ_W6BM3tg55SDZdVbHsPb10BCX3W3fYZHv4Xw4pdW6fRGCy8slM_lW5r8Bg91kF-2hW1y36L18wZyG3W6pJG3Y5XVlXjW4vsZHs4b2stbN7rbthn4nwCrW1XM0cX4YG2T4W6Rk-kq75_CdzW7cqVZx2NTTvCW6HPgdf3mF_rxV9TVtd2lyZL4W8YKF9h3hHpxYW20Wsg81yknNpW1dYQ4y6g8q_0W2rkL__2qrBLdN774pLxVh5YhW179RlC1lyT1CW4S9pmy5fW3mbW8cTBV46r75ckW38NDtf12RYsPW4qJ0kx4BBjCdW34CkJ11z7CK-W47S6kl93WJ8sW4Fw4964FmQfvf4svdCC04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">study</a> reveals that people used Copilot differently late at night on their phones than during the workday on their laptops. Conversations that focused on productivity and career were more likely during the day and on desktop devices, and health, gaming, and philosophical questions dominated non-work conversations. As 2025 went on, more users asked the AI agent for personal advice.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">How it works:</span> Researchers <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHrx3qgz0W95jsWP6lZ3kSW3_s33w6nj_V5W1NsHyQ16gdWYW7FsRTD7d0WydW41QJ8J4fRfk9N3ptYklrjwq3N7MRQD3PWPy2W3JgDy-53HBKQVgd_z97LDn6PW4DXzYT1678thW5qWlPB1HH1Y-W3SRzGS1HjyqxW6z-Lw480bPc0W2F-yv98JmTHsW5kJBSg2WqnSNT8v7175Rdf-W28ngBM3zbtNsW63MJKw8fvpQbW8dd7xm2Lh7bhW8hhhqK1DPtNlW4zlr2-8lTDHFVV-cVW1LkLKdW1gQhPL5wLzPCVpcjYG1pwDhYN2tVwmsZQ57pW4_z5h26QP7_gW4TgYVy6gml4gVgK4hl6fqQfzVYxSvP47T8d5W4P5ZRN2B-s9VW4kgqff6nkTWVf6GCBcF04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">analyzed</a> anonymized summaries of 37.5 million Copilot conversations between January and September 2025 to study how customers used the system, making this the largest study of its kind to date. The authors conclude that AI has become more socially integrated, as users employ it in aspects of their lives beyond work.</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">The authors examined a random sample of Copilot conversations by paid and unpaid users, excluding commercial, enterprise, and education accounts. Each conversation included timestamps and device type. The authors used AI tools to summarize roughly 144,000 conversations daily. They built classifiers to assign each summary a topic (like “technology”) and intent (like “seeking advice”), identifying about 300 topic-intent pairs.</li>\n<li aria-level=\"1\">The study ranks the frequency of topics and intents by time of year, time of day, and device type. The top 5 topics in order were (i) technology, (ii) work and career, (iii) health and fitness, (iv) language learning and translation, and (v) society, culture, and history. The top intents were (i) searching, (ii) seeking advice, (iii) creating, (iv) learning, and (v) technical support.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Analysis:</span> Topics and intents differed depending on device used, time of day, and time of year.</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">Users were much more likely to discuss health and fitness on mobile devices than desktops. Seeking advice about personal matters spiked near Valentine’s Day. Philosophical questions became more common later at night, while entertainment-related conversations plummeted during the workday.</li>\n<li aria-level=\"1\">As the year progressed, topics and intents became less focused on work and technology and drifted towards social and personal matters. This shift suggested that the user base became both larger and less technical and/or users began using AI for both personal and professional matters.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Behind the news:</span> Microsoft’s report follows similar studies by some of its AI rivals. </p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">In September 2025, OpenAI and Harvard released a <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHpM5nR3bW5BWr2F6lZ3lRW37hpZn1mJq4NN3YRn_qB0gp0W3K-g-h8lhHdCW1GpTqN7FyWflW15dtnt4r-qWQW8-rrXJ8XM6-GW5gkrNP3G2dX1V2y34w54Wwy8W60QD2_9b5kP5W4Mt5W21N4q9tW7Km6kG2pjpZPW9f0-GR6_C3RLVYd6LC7LGhRvW7YmfCp1_xcTSW4JNMr33bsLpjW7mdglL569p2xMssSLKZ_Wc7W65ljLW3Wrl1hW84XCly7hSxHZW87CF0z4lwQxMW6v2Nw-7xR94tW6DGHVQ2jn857W7FPV_q76RMzsW99mxs_4fGXMvW8cg52j8d1hDpW5MvrJn5RcXcGW83glV05l-5mRW4r2KTj1xCWnNW49Bzmy7XqtMWW2Y3qF47bRPXDW6nHx592MpnKpW73y5NL1LcMz7W49HGfZ28RJFtW61Cpqp5C0Rjqf1CKkLY04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">study of ChatGPT use</a> from 2022 to 2025. It showed that 30 percent of uses were related to work while 70 percent were related to non-work activities. In addition, the gender gap among users among users shrank steadily during that period.</li>\n<li aria-level=\"1\">In January 2025, <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHps5nR3bW50kH_H6lZ3n2W3FZwH72gQLPmMDFTCMKTk4JW8Sp0Dm8m1h3NW8RH7W_3V1lL5W4lHG7X5MX36_W23hmkW5t954HN1DfYFhV0TkBVh4gP329XMV9W82F4JW3c8qCmW1wfyqy4z-kVtW44zxzn8cl1ZhW4by7Yh8_szdzW4hP3M07VxSjHW4nQczv18c48BW1ss5yQ59yckrW760lpp6WW-5lW2lhgnj292JLdW6zZ1yZ7Bhg3WMJRhrdCk4LdW7dzMVK1Q-mkKW32N-Mb3SXM5fW43YmQ560pSdtW8x1HFj8L6k4xW2dSKHF4pBNSDMGr_3q7z1ZyW5GPGcj98lTFNW1vDbcF9gjmwdW735N-b2VC6QtW4hk2sF2XNm9SVT9Z235BP9zCW3RPKBt1xs5zsW4DhHvy5JMT3Ff4FzF6404\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Anthropic’s study of Claude</a> showed that the model’s user base focused on work, especially software development and text communications. A small but growing number of users engaged in games like Dungeons & Dragons and sexual roleplay (despite prohibition of that use by Claude’s terms of service).</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Why it matters:</span> The authors argue that the AI community may need to rethink chatbot design altogether. If users treat chatbots differently on mobile and desktop devices, AI builders would do well to design their systems to suit the devices that will deliver them. Application design is one way to accomplish this, but system prompts may be another. Desktop chatbots and agents can respond with more information-dense answers, guiding users to execute tasks, while mobile agents can offer shorter, more empathetic responses.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">We’re thinking:</span> Studies of chatbot usage conducted by different companies show different results. Perhaps each company’s users treat AI differently, so the results of any given study may not apply generally. That said, the Microsoft study suggests that the device used and the time when it’s used can have a big impact on what users want — important considerations for designing any application.</p>\n<p style=\"line-height:150%\"> </p></div></div>\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" width=\"100%\"><tbody><tr><td style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old22\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"none\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" style=\"vertical-align:top\" width=\"100%\">\n <tbody><tr>\n <td align=\"center\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word\">\n <p style=\"font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto\"></p>\n \n </td>\n </tr>\n</tbody></table></div></td></tr></tbody></table>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_17241662955853\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;word-break:break-word;text-align:center;padding:5px 0px 10px;font-size:0px\">\n <img alt=\"Diagrams comparing LongCoT and Delethink environments show reasoning processes and context management.\" src=\"https://info.deeplearning.ai/hs-fs/hubfs/MARKOVIAN.png?width=1200&upscale=true&name=MARKOVIAN.png\" style=\"outline:none;text-decoration:none;max-width:100%;font-size:16px\" width=\"600\" align=\"middle\" class=\"m_-6958527957600471347stretch-on-mobile\">\n </td>\n </tr>\n </tbody>\n</table></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_module_17370395015403\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_module_17370395015403_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px;font-weight:bold\">More Affordable Reasoning</h1>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\">One way to improve a reasoning model’s performance is to let it produce a longer chain of thought. However, attending to ever-longer contexts can become expensive, and making that attention more efficient requires changes to a model’s architecture. Researchers proposed a way to limit the cost of processing long chains of thought with just a bit of training.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">What’s new:</span> <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nKW8SXyhJ7DQcqSW8nHYDW27ZnkMW7-qqxL1DZ4LXW99kBsN2HH6cXN3hSL0PzN9_GW88SD2t62Xn-NW7jC2pZ5GwK4CW5NRxGv8YSxJpW58GnTd4-W-PbW4VG01n68vhq3W3c4Hvz3trwNJN5cdb0q5KsgxW6q-32N7fK-0lV4JBTF5qrtGGW9jgDm36dLyBMW91y9R58pXr96W69J70p7dz_LvVbq9_b5gbqjmW12FCCy3YdKTPMWv2gLQgnBKW2Qvm1-1VWPlsW8731bg6JnDL0N3Ry2sX9pj82W2l-hmn1M1NwDf8Qx6z404\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Delethink</a> is a reinforcement learning (RL) method that trains large language models to periodically truncate reasoning tokens to a fixed maximum number. The authors include Milad Aghajohari, Kamran Chitsaz, Amirhossein Kazemnejad, and colleagues at Mila, Microsoft, McGill University, ServiceNow Research, Polytechnique Montréal, and Université de Montréal.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Key insight:</span> Reasoning tokens typically accumulate within a large language model’s context window, where they consume quadratically more computation as the contents of the window expand. One way to counter this effect is to train the model to reason within a maximum context window size. In effect, as a model is reasoning, it can learn to replace its chain of thought periodically with its latest “thoughts” and then continue.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">How it works:</span> The authors fine-tuned R1-Distill 1.5B, a large language model, on math problems in the <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHq25nR3bW69t95C6lZ3mTW80Q7xZ4QXs__N1lxhWpcckqyW8wVFFy6hzJcVW7lSDK17bJLCDW4lhZLT5Sd68CW9h4NJR60THQ-W5k4NRL8Q42SYW80l_MJ80xtntW6wHP2T7R1_QvM3MrcZJQBj4W7sKMf89hjZJZN7mGhVXYlC5JW75Xhct8CB4X6W48SvnV7pRGdlW3LDlbC5mZnl1W1lM38N4LctGlW4Kvmxd7_MBlFW2D9djG45RMj6V3PgnB349h89VDJ0N86CN-X-W91qhB675mcPDW3Xv7FR3zltJCN6mtqCgyZ0QYW1h0rpD6KVCZYW2lzJK27Fp93jW3rtB8P6p3RQbVT5mqm8LrDz8W4-kgsq6p-6fTW3MytCj3R3WLpW38hspr2T_sqRVXXhZ_3fxpH4W1s9Xp_36n6YCW9681k46Q2Cn-W804H7f3-t3N3W4ld37c4JpVXDW6dp7BY8lg1Z5f3ztR6-04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">DeepScaleR dataset</a>. They used a modified version of the reinforcement learning algorithm GRPO that trained the model to reason in 4,000-token chunks:</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">Given a math problem, the model generated a chain of thought until it had either finished or filled the model’s context window with 8,000 tokens.</li>\n<li aria-level=\"1\">If it didn’t finish its chain of thought, the authors replaced the context with the original query plus the last 4,000 tokens. Then the model continued to generate its chain of thought until it had either finished or the context window once again held 8,000 tokens.</li>\n<li aria-level=\"1\">They repeated this process until the model had either finished its chain of thought or produced 24,000 reasoning tokens.</li>\n<li aria-level=\"1\">Then the model attempted to solve the problem, receiving a reward for a correct solution.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Results:</span> The authors compared their R1-Distill 1.5B models to the same model after fine-tuning on the same 24,000-token reasoning budget via using GRPO. They tested the models on reasoning budgets of 24,000, 96,000, and 128,000 tokens.</p>\n<ul style=\"line-height:150%\">\n<li aria-level=\"1\">With a budget of 24,000 tokens, their model matched or surpassed the baseline on all 3 math benchmarks tested. For example, on AIME 2025, Delethink (31 percent accuracy) outperformed the baseline (29 percent accuracy).</li>\n<li aria-level=\"1\">Their model’s performance continued to improve as the authors increased the reasoning budget, while the baseline achieved much smaller gains. For instance, with a budget of 128,000 tokens, their model achieved 35 percent accuracy, while the baseline achieved 30 percent accuracy.</li>\n<li aria-level=\"1\">The authors estimated that training their model with a 96,000-token reasoning budget would cost 7 H100-months, while the baseline would require 27 H100-months.</li>\n</ul>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">Why it matters:</span> This work eases the quadratic compute barrier that can make extremely long reasoning computationally infeasible. While other methods, like linear attention, achieve the same result by changing the attention mechanism, Delethink restructures the reasoning process to limit processing regardless of a model’s attention mechanism. It opens a path to reason efficiently over longer contexts without requiring new model architectures.</p>\n<p style=\"line-height:150%\"> </p>\n<p style=\"line-height:150%\"><span style=\"font-weight:bold\">We’re thinking:</span> As the authors mention, most LLMs are pretrained using relatively short contexts. For example, Llama 3 models started pretraining with examples of 8,000 tokens. This may have made them good at processing inputs around 8,000 tokens long. That is to say, Delethink’s performance may have been helped by the fact that LLMs tend to be pretrained on short-context tasks.</p>\n<p style=\"line-height:150%\"> </p></div></div>\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" width=\"100%\"><tbody><tr><td style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old25\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"none\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" style=\"vertical-align:top\" width=\"100%\">\n <tbody><tr>\n <td align=\"center\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word\">\n <p style=\"font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto\"></p>\n \n </td>\n </tr>\n</tbody></table></div></td></tr></tbody></table>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old29\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old29_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><h1 style=\"margin:0;font-family:Helvetica,Arial,sans-serif;line-height:125%;font-size:32px\">Work With Andrew Ng</h1>\n<p style=\"line-height:125%\"> </p>\n<p style=\"line-height:150%\">Join the teams that are bringing AI to the world! Check out job openings at <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3p5N65swpFxLdcBW8s0KY56mqGLPW54Gqb655n5yLW2VlKw93KfzxyVsN0Xb6JKmjBW58N5lm172PBQW1pyv7s2nLPmJW425jTD7Kg0ltW6L4phx2-p_-fW4f56hZ8KXv_QW92xSwf4_ytncW4592Ww3PryZ1W8573kw11T8gtW11NgFR1dbB96W3YbCdl4xQrpNN81SyzgkdP_TW3X5ZmK4ZtpbnW8SFYXH5z_h9FW56-zy35zYthhW7Z8hQ14HdSpGW86N-fM4Fr3J2W6sVzYm75ZmQDW1yYb9B41ckycVXn2j82_vpHcf6T8cZl04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">DeepLearning.AI</a>, <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3lbW49jm1g3GPY-6N6cnvR7rjkr-W8VgP9N5JN8qsW6yShD98ZcZqKVrKLqv3dPMCpW718j8y5JdRKMW2tsL9K4nVRPnW686Cl16p2Bc4V9V3rl94vpRvVm48ql1qZpmxW68CxMZ5CfhzDV5KS6l8P2GyXVWcLT93KYJZcW4cs8CQ4dWkV_W8-rzXF4LCJjWV7Ltx08PgvF6W3ZKBYn7LjkykW4w4bvB4b9dgpW5q3QZW987twRF903KzsnMM2W7r5RZ12C1vQWW6bZJhw9htKGMW8lFfhm2DbYx_W7-NxNX7GdCfMf3Z3kkT04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">AI Fund</a>, and <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3qgz0W7lCdLW6lZ3nTW1lpkY37j3bCYW33RpBy4V29pwW98gl3t4BfKvGW4zYqFT5Qjxl6N1RXgGSVm1WrW7xP5T35MzQ_LW8rWqPX5ZW-dKW4DVSpQ1S4YbXW392rhC2-v-X3W6Fs7xg7RtQVdW9llkVF7-jWgmW2Py9Qd71P794W77Js7c6wGQy9W1mDLHH4MFp59W7FbMRR989XsQW97sLMj6YmLRbW3GXV_R3XnjFyW2qLwG01d0fwXVbzPfm2hWP9hW2Y8Htq6wMdHkW8PqLSG2S1-NjW8G4rfc5dtp1SW8jKk_Z44jcmMW6WBJ4K1Z0S1_dZvf6804\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">Landing AI</a>.</p>\n<p style=\"line-height:150%\"> </p></div></div>\n <table role=\"presentation\" cellpadding=\"0\" cellspacing=\"0\" width=\"100%\"><tbody><tr><td style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;padding:0px 0px 10px\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old30\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"none\" border=\"0\" cellpadding=\"0\" cellspacing=\"0\" style=\"vertical-align:top\" width=\"100%\">\n <tbody><tr>\n <td align=\"center\" style=\"font-family:Helvetica,Arial,sans-serif;color:#3b3b3b;font-size:0;word-break:break-word\">\n <p style=\"font-size:1px;border-top:1px solid #000000;width:100%;margin:0 auto\"></p>\n \n </td>\n </tr>\n</tbody></table></div></td></tr></tbody></table>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old31\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><div id=\"m_-6958527957600471347hs_cos_wrapper_hs_email_body_old31_\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><p style=\"line-height:125%;font-size:14px;text-align:center\" align=\"center\">Subscribe and view previous issues <a href=\"https://info.deeplearning.ai/e3t/Ctc/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsN6wcHqF3lYM-W7lCdLW6lZ3prW7Qb8g-2wx0YNVy8LrQ4CN6dbW3MHBx77698fhW8QshbW4N0vzqN4Prl3f180wZW1MF8S644trRSW1SwZL88hTcWfW2t4t953T985BW7vH8P03bfd35W86y-3F1K69nqW8Tg_6s4rlkY0W5_7hHx7fv5-RW1rtZ032W9C_xW3-NFd_79zfRNW1WgY_z5R8lQzW7YzSr-2J5t3TW59mPLM72YyqMW2XhG3v2PBjKcN2Bj7GGwPWv6MYndxzMw1qGW1N2Vmb1Qr12jW51_Yxt7LG-W1W1tjyPX7mh_KhW7LJPvS60H8jDf1gjZhP04\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">here</a>.</p>\n<p style=\"line-height:125%;font-size:14px;text-align:center\" align=\"center\"> </p>\n<p style=\"line-height:125%;font-size:14px;text-align:center\" align=\"center\">Thoughts, suggestions, feedback? Please send to <a href=\"mailto:thebatch@deeplearning.ai?subject=RE:%20Thoughts,%20Suggestions,%20feedback\" rel=\"noopener\" style=\"color:#237b94\" target=\"_blank\">thebatch@deeplearning.ai</a>. Avoid our newsletter ending up in your spam folder by adding our email address to your contacts list.</p>\n<p style=\"line-height:125%;font-size:14px;text-align:center\" align=\"center\"> </p></div></div>\n<div id=\"m_-6958527957600471347hs_cos_wrapper_footer\" style=\"color:inherit;font-size:inherit;line-height:inherit\"><table role=\"presentation\" width=\"100%\" cellpadding=\"0\" cellspacing=\"0\" style=\"font-family:Arial,sans-serif;font-size:12px;line-height:135%;color:#23496d;margin-bottom:0;padding:0\">\n <tbody>\n <tr>\n <td align=\"center\" valign=\"top\" style=\"font-family:Helvetica,Arial,sans-serif;font-size:18px;color:#3b3b3b;word-break:break-word;text-align:center;margin-bottom:0;line-height:135%;padding:10px 20px\">\n <p style=\"font-family:Helvetica,Arial,sans-serif;font-size:14px;font-weight:normal;text-decoration:none;font-style:normal;color:#3b3b3b\">\n DeepLearning.AI, 400 Castro St., Suite 600, Mountain View, CA 94041, United States\n </p>\n <p>\n <a href=\"https://info.deeplearning.ai/hs/preferences-center/en/direct?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159\" style=\"font-family:Helvetica,Arial,sans-serif;font-size:10px;color:#3b3b3b;font-weight:normal;text-decoration:underline;font-style:normal\" target=\"_blank\">Unsubscribe</a>\n <a href=\"https://info.deeplearning.ai/hs/preferences-center/en/page?data=W2nXS-N30h-BnW1Vq2n33dynlvW1_l4nf2p5qhgW2vXFx732F24JW1VhnN345Cb7BW3_JQdS1VpCSdW34DsRl22_R6JW3XJSW74kdCl-W2nMM4g2MLt2wW4cgBk73VyFh_W3y-MJg4ktYC8W3QQqP24t4pdfW3GJy0m3zdxhQW34kJP82YfWz8W22W0gK3VHRf6W2t2-cH4txxNCW2HL53d4mC36LW3Q_Tct4kLJqzW1_jWzL3T5XHsW1N56M53zd1NQW36m6dc22V8fyW4pkfTk4md19xW2zz8HP3P1hpVW2HCRsF4tpPrWW4tBmlm3XJ0T8W213m8X4mJqSsW2r2P6Q2ThyN_W22V6dg3DXXGpW3f-P-31S1nQdW2z_QlL3JQbTHW1QxgmV41W2k0W2MXZjZ41KQRpW3Y0SSG3XsyGLW1VwqNv2WKhQ3W3gy8XF1QtLB_W1LC44Y4fpRrpW3gxm1r41z9tzW34kcF22sN5XrW34HqlZ4pxZg5W4chp3d1NcsKgW4cxJVH2Hy3j-W32B8xs49QQMFW3hYQNM4rzLLfW24WG0y4fgLc7W32gy6j2vS4_mW4twDfZ32q1S5W30n9wf47pyVjW2zyLT13M37MwW2PP9nH1L95CwW2WrHjf3ZFFzTW3gpDvz3d1W83W36jfh93QWp4tW1Z5n2-32j_4FW3LHN1Q4t7GrPW2t2tWd49gmGCW3c-8ph3XzmV0W38CBBQ4rnm-90&utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&utm_content=397569578&_hsenc=p2ANqtz--55DTHo2rOVmyG2HV2HZhOcWdUxppSKvfRhZf7zWT6kQMZF_-nIQ6J-uhYAD2CWrbiadR3lAZOLtUqVuPXMsQ58o0F00OWowHnx_1W8G1hlGW7SAA&_hsmi=397572159\" style=\"font-family:Helvetica,Arial,sans-serif;font-size:10px;color:#3b3b3b;font-weight:normal;text-decoration:underline;font-style:normal\" target=\"_blank\">Manage preferences</a>\n </p>\n </td>\n </tr>\n </tbody>\n</table></div>\n</div>\n\n\n \n </div>\n \n \n \n </div></div>\n </td>\n </tr>\n </tbody></table>\n </div>\n \n<img src=\"https://info.deeplearning.ai/e3t/Cto/LX+113/cJhC404/MWqCl5gc-gvW3CQK_18qz2D6W1m0cDV5J4hrsW6wcH9r3Cy50v122\" alt=\"\" width=\"1\" height=\"1\" border=\"0\" style=\"display:none!important;min-height:1px!important;width:1px!important;border-width:0!important;margin-top:0!important;margin-bottom:0!important;margin-right:0!important;margin-left:0!important;padding-top:0!important;padding-bottom:0!important;padding-right:0!important;padding-left:0!important\"></div></div></div><div><br clear=\"all\"></div><div><br></div><span class=\"gmail_signature_prefix\">-- </span><br><div dir=\"ltr\" class=\"gmail_signature\" data-smartmail=\"gmail_signature\"><div dir=\"ltr\"><div><div dir=\"ltr\"><div>Markus Edgar Hormeß<br>Founding Partner<br></div><div><br></div><div>WorkPlayExperience GbR<br>Oedenberger Weg 14<br>D90571 Schwaig b. Nürnberg<br>Germany<br><br>+49 151 24101032<br><a href=\"mailto:markus@workplayexperience.com\" target=\"_blank\">markus@workplayexperience.com</a><br><a href=\"http://www.workplayexperience.com\" target=\"_blank\">www.workplayexperience.com</a></div></div></div></div></div></div>\n", "subject": "Fwd: LLMs Go To Confession, Automated Scientific Research, What Copilot Users Want, Reasoning For Less", "inbound_domain": "inbox.workplayexperience.com", "attachments": [], "channel": "email", "event_id": "143eaa4b-f8d4-4768-ae85-d1f0587989ef", "transaction_id": null }